Claude Code, Codex and pi each append every prompt, answer, command and edit to a transcript on disk. ravnori indexes those files directly, so capture never waits on an LLM, a worker, or a queue, and one memory is shared by every agent you run.
Memory inside one agent is becoming a free built-in (Claude Code’s auto memory, Codex’s memories), and line-by-line AI attribution has a funded owner (git-ai, and Cursor’s open Agent Trace format). What nobody owns is the layer underneath: one complete, local, cross-agent record that loses nothing, cites its evidence and knows whether its code survived. A dozen tools each re-read the same transcripts with their own fragile parser. ravnori makes that layer the product, under the AGPL-3.0, and memory becomes the first app on it.
Rules: unique (only what the record makes possible), measured (every claim has a number, and recall changes pass the gate), easy (one command to install, nothing to configure), and fast (hooks under 300 ms, one binary).
rvn ask, rvn trace.On this machine, claude-mem stored nothing for 157 of 493 sessions (32%) in the 30 days before the switch, and left 883 sessions stuck as active. Its observer also ran 1,451 hidden Claude Code sessions to summarise the real ones.
claude-mem has since moved partly off this path. Since 7 September it has stored raw tool calls (input and output, except private prompts and its own memory tools) in a tool_uses table before any model call, which this page missed when it was written; since v13.30 (4 October) its hooks write each event to a spool on disk; and since v13.36 (12 October) a provider that says no pauses the observer instead of dropping the work. For hosts with hooks, capture still needs the hook to run, and the transcript watches its installers set up start at the end of existing files. Not re-measured on this machine.
A crash, a skipped hook or a model outage delays capture; it cannot lose it. ravnori re-read all 1.45 GB of this machine's transcripts from zero and stored 0 duplicates. It also recovered 30 of 30 Codex sessions claude-mem had missed.
Thirteen tools checked on 28 September 2026 against their own READMEs, docs and the GitHub API. Cells marked partial carry a note. Where a fact could only be inferred, the note says so. Cursor's Memories feature is left out: it is missing from Cursor's current docs and appears to have been removed in 2.1.
| Tool | Capture | Saved without an LLM call | Agents | One store across agents | Recovers from old transcripts | Search | Runtime | Licence · cost |
|---|---|---|---|---|---|---|---|---|
| ravnoriunreleased · this machine | Automaticindexes transcript files; hooks + 30 s watcher | YesLLM distils later on its own thread and catches up on missed sessions; any OpenAI-compatible endpoint, or the user’s own Claude Code or Codex sign-in, with model fallback | Claude Code (all profiles), Codex, pi | Yesplus per-prompt “meanwhile” updates | YesClaude Code, Codex, pi transcripts; claude-mem DB import | HybridFTS5 BM25 + local MiniLM vectors in process; meaning gate on recall | One binaryRust, SQLite; nothing else to run; background service under systemd (Linux, WSL) or launchd (macOS) | AGPL-3.0, commercial licence available · free |
| claude-mem101,063★ · v13.39, rechecked 16 Oct | Automaticlifecycle hooks spooled to disk; transcript watch for hosts without hooks (DeepSeek Harness, Grok Bot) and, opt-in, Codex subagents | Raw calls onlyraw tool calls stored at capture (since 7 Sep); the memory itself waits on the model | 10+ hosts incl. Claude Code, Codex, Cursor, OpenCode, pi | Yesinferred | Opt-ina transcript watch you configure yourself reads existing files; the ones its installers set up start at the end; import requested in issue #3763, still open | HybridFTS5 + Chroma vectors | HeavyNode 20+, Bun, Python/uv for Chroma | Apache-2.0 · Pro $30, Heavy $70, Max $100/mo |
| agentmemory28,937★ | Automatic12 hooks in Claude Code, 6 in Codex | YesLLM compression optional | 9 incl. Claude Code, Codex, Cursor, Gemini CLI, pi | Yes | Claude Codeimport-jsonl for Claude Code transcripts | BM25 → hybridhybrid once embeddings are configured | Node + engineNode 20+ and the iii-engine binary | Apache-2.0 · free |
| claude-mem-lite61★ | Automatic7 hooks plus /lesson, /bug | YesLLM batched per episode, saved even if it fails | Claude Code only | No | YesClaude Code transcripts | Full-textFTS5 BM25 with query expansion | NodeNode 22+, SQLite | MIT · free |
| Mem066,085★ | Hooks + toolsplugin hooks at session boundaries; MCP tools the agent calls | Optionalfact extraction by default; infer=False skips it | Claude Code, Codex, Cursor, Windsurf, OpenCode | Yesper account, inferred | Noinferred | Hybridsemantic + BM25 + entity | ServiceDocker self-host or cloud API | Apache-2.0 · cloud tiers |
| Supermemory30,949★ | Automaticplugin capture | Noprocessed server-side, inferred | Claude Code, Cursor, Codex, OpenCode +2 | Yesper account, inferred | Unknown | Vector + graphinferred | Cloud-firstself-host possible | MIT · credits |
| Hindsight37,059★ | API callsretain API and integrations | Noinferred | Claude Code, Codex, Cursor, Copilot + | Yesinferred | Unknown | Hybridsemantic + BM25 + graph + temporal | HeavyDocker, Postgres, LLM key | MIT · paid cloud |
| Graphiti / Zep31,211★ | Tool callsagent calls MCP or app adds episodes | Noentities extracted by an LLM | Any MCP client | Yesinferred | Noinferred | Hybridsemantic + BM25 + graph | HeavyPython, Neo4j/FalkorDB, OpenAI key | Apache-2.0 · Zep Cloud paid |
| Basic Memory4,050★ | Agent writeswrite_note over MCP | Yesinferred | Claude, Codex, Cursor, any MCP client | Yes | Chat exportsClaude.ai / ChatGPT exports, not CLI transcripts | Hybridfull-text + FastEmbed | Pythonvia uv | AGPL-3.0 · cloud $15/mo |
| ByteRover4,966★ · ex-Cipher | Curate + reviewuser or agent curates, then approves | Noinferred | 22+ agents | Yes | Unknown | Unknown | Nodebundled binary or Node 20+, LLM provider | Elastic 2.0 (not OSI) · Pro $15/mo |
| Letta Code3,434★ | Self-editingagent rewrites its own memory blocks | Noinferred | Its own harness only | Noinferred | Unknown | Unknownmessage search | Node + serverLetta Cloud by default | Apache-2.0 · cloud plans |
| Claude Code built-inauto memory | Model decideswrites Markdown during the session | Yesno separate call, inferred | Claude Code | No | Noinferred | Nonefirst 200 lines / 25 KB loaded | Built in | Included |
| Codex memoriesoff by default | Automaticbackground summaries once sessions go idle | No | Codex app, CLI, IDE | No | Yesbuilt from past sessions | Nonesummary injected, ~5k tokens | Built in | Included |
rvn doctor reports lag and lost bytes instead of staying quiet.get_observations prints the evidence and flags events that changed after the memory was written. claude-mem keeps the raw tool calls and points each batch of them at the first memory it produced, a link it documents as not one-to-one; no other tool we checked keeps the source behind a summary.rvn doctor shows the backlog. An installed agent that is not connected, or whose Codex hooks are not trusted, is flagged too, with a Connect button in the viewer.rvn install picks one when nothing is set. An ordered model chain falls back when a model hits quota; tested live when gemini-3.5-flash-lite returned HTTP 503. Distillation runs after capture, so a failed or slow model never loses a session.BeforeAgent hook that adds context); Cursor’s prompt hook can block but not add text, and OpenCode keeps sessions in SQLite. Added when someone needs one.The seven essentials, ordered by value to a daily user divided by effort. Each passed two code-review rounds by GPT-6 Astra, Gemini Pro Agent and Claude Fable 5.1, which found and forced fixes for real bugs before release. 60 tests cover them.
The watch service takes a verified snapshot every night and keeps seven. Each snapshot is reopened and checked (integrity, schema, row counts, SHA-256) before it counts. rvn restore proves a snapshot on a scratch copy; --apply copies it into the live database through SQLite’s online-backup API.Council found a restore that could drop a concurrent write and one that could wait forever; both fixed and tested. Restore now gives up after 30 s with nothing changed.
Summaries cite the transcript events behind each claim; citations outside what the model was shown are dropped. 94 links so far across 47 new memories. Imported claude-mem memories say plainly that no source was kept.Council found a citation that could slip in when a long turn was cut; the allowed list now comes from the text actually sent.
At the end of each turn ravnori records branch, HEAD, unpushed commits and uncommitted files. The next session in any agent opens with a “Working tree” section, labelled with its age.Runs detached with one 3 s budget for all git calls, so a slow repository never delays the agent.
Each prompt adds up to five memories that match it, never repeated in a session. A memory needs two of the prompt’s words and the prompt at least one rare word; personal-detail memories are never injected automatically. (An earlier version of this page said the rare word had to be in the title; that check never existed, and adding it did not help.)Measured with the recall test: hit@5 87%, hit@1 55%, MRR 0.68, p95 about 110 ms.
rvn forget removes memories, events, sessions or projects and leaves a marker that ingest, import and summarisation all respect. rvn remember pins facts every agent sees; rvn pins lists them.Council found three ways forgotten data could return (re-import, in-flight summaries, prompts); all closed and tested.
One set of checks feeds rvn doctor, the viewer and session start, where Claude Code and Codex show alerts to the user.Expensive checks moved into the watch service; session start stays around 30 ms with no false alarms.
rvn export writes everything as JSONL with a meta header. rvn eval builds known-item questions from real memories and reports hit rates and latency; the set stays private on the machine.Used immediately: it picked the ranking weights and rejected a change that looked good but was not.
Six more steps, each reviewed by the council (GPT-6 Astra, GPT-6 Sol on operations, Claude Fable 5.1) across rounds 12 to 25. Every number here comes from a test set, and the last two steps were chosen on data the tuning never saw.
Memories get vectors from a local model the watch service keeps loaded; hooks ask it over localhost and fall back to keywords within 300 ms. Vectors are tied to the exact text and model files they came from, so an edited memory or a new model never meets a stale vector.Council rounds 12–15 closed a stale-vector race, a revision mix-up and a port the hooks could not find.
MCP search used to need every word, so a plain question found nothing (0% on the test questions). It now takes all-word matches, then any-word matches, fused with the nearest memories by meaning; the viewer shows one ranked list with a words / meaning badge on each card.Through MCP search: top 5 went from 0% to 85% of test questions. Pages are stable and one limit spans memories and transcript events.
Keyword hits that share words but not meaning are dropped before they reach the agent. Search applies a softer version to any-word matches only.Prompts with no answer that still recalled something: 9 of 10 down to 1.
rvn eval --build-real samples real prompts, split by session into a tuning half and a test half. Each is replayed as of when it was typed; --judge has the distillation models mark what recall showed, with a second judge and its agreement on request.First labels from file overlap were dropped after reading results: they missed most relevant memories. Judges agree on 425 of 541 memories (kappa 0.57).
Recall leaves out memories distilled from the session that is asking; the agent already has that conversation in context.Found by the judged eval: 43% of judged memories were the session’s own, and they took 55% of the “helpful” marks.
Seven local models screened on the tuning half; all-MiniLM-L6-v2 separated helpful from unhelpful candidates best (AUC 0.731 against 0.657; bge-small at 0.711 is within noise). Each model gets its own thresholds.Held-out, run once: helpful memories shown 149 → 179 (luna) and 100 → 141 (haiku), unhelpful 95 → 93 and 144 → 131. Trade: one test set lost two first places.
Take a verified backup in the viewer, download it, and on a new machine upload it, see it next to what that machine holds, then replace. The backup carries the settings (never keys); they are off unless ticked, and any change to where prompts go is flagged.Council rounds 26–29 found a trigger a crafted backup could smuggle in, a write that could be lost mid-restore and a missing upload deadline; all fixed and tested, including a full two-machine move.
rvn eval --gate runs the installed ravnori and the candidate build (or candidate settings) on their own copies of one snapshot, graded by the same judge, and refuses the change if recall got worse: 17 checks across known, vague and judged real prompts, MCP search, personal details listed unasked, and hook latency timed in real processes.Checked: an identical build passes 17 of 17; switching the relevance gate off fails 5; a build without ONNX fails 4; a slow settings file fails on latency (631 ms). Council rounds 31–33.
When Claude Code first reads or edits a file in a session, ravnori adds up to three past memories about that file, each with how the file changed since in git (commits, uncommitted edits, lines added and removed). Any agent can ask with recall_file(path); get_observations flags memories whose files changed. About 40–170 ms, once per file.Measured first on 172 real edits: 44% had memories about the file; 74–79% judged helpful with the file named, 51% with it hidden, so about as useful as prompt recall and almost never the same memories (4 of 198). Council rounds 34–37 fixed cross-package and node_modules matches, pattern-like file names, stalled git and binary files.
Three cross-encoders scored each prompt’s ten keyword candidates on the judged real-prompt set. None beat the MiniLM gate by more than noise (AUC 0.702–0.745 against 0.713) and each took 0.9–3.4 s per prompt against a 300 ms budget.Only reordering the keyword top ten was tested; a wider candidate pool and reranking MCP search, where seconds are acceptable, remain open. rvn eval --rerank re-runs the comparison.
Six changes, reviewed by the council (GPT-6 Astra, GPT-6 Sol, Claude Fable 5.1) in rounds 41 to 44, and one idea measured and dropped.
The watcher distilled only the 5 newest sessions of the last 2 days; the rest never became memories (862 in 30 days) while doctor said OK. It now distils on its own thread, then backfills the last 7 days oldest first under a daily request budget, and warns when sessions will age out. rvn distill --since-days 30 --max-calls 200 catches up older ones by hand.Council rounds 41–42. First catch-up batch: 200 calls, 539 memories, no errors.
claude-mem stopped importing, so every new memory is distilled by ravnori, yet only 4 of the 70 judged real prompts could see one. rvn eval --build-recent samples prompts from the last 30 days in projects that already held ravnori memories (at most 6 per session, half per project); the gate now judges them too, with the same limits. On its first run ravnori memories reached the top five for 16 of 57 prompts, and 34 of the 38 shown were judged helpful against 74 of 121 imported ones.57 prompts per half from 30 sessions; a baseline built before the set skips its checks.
A memory about a file said whether the file changed since, which in an active repository is nearly always: “3 of 3 predate changes” on every touch. It now says whether the lines its session left are still there, partly (how many) or gone, reading each edit back from the transcript record it came from. Of 354 recent memory-file pairs that could be told, 268 still had their lines, 77 partly, 9 not. Only memories whose edits changed count toward the warning.Council round 43: exact tool call and record identity, edits applied in order within a chunk, whole-line matches, bounded reads. File hook p95 78 ms against 75 before.
24 of 2,482 sessions were driven by another agent’s brief (a review council, test runs) yet made 17% of a week’s distilled memories, and 9% of the memories shown to real sessions came from them. Prompts matching scripted_sessions patterns now mark a session: it is not distilled or offered memories, and its memories stay out of recall and the session-start context. Explicit search still finds them, labelled. Nothing is deleted.Council round 44: the last-summary line, timeline neighbours, an atomic rescan when patterns change, and search kept as a pull.
Each gate run copies the database and removes the copy when it ends. A run that was killed left a 423 MB copy behind; the next run now removes copies older than six hours.
A title rule that puts the component, file or decision first was compared on 30 chunks the live prompt had distilled, with a question per chunk written from its digest by another model. Rewriting only the titles, contents unchanged: top 5 fell from 22 to 18 of 30 and titles-only choices to open from 19 to 16. It helped where the old title named nothing (a phase number followed by “planning completed”) and hurt where a path or id replaced the words people ask in (a title about merge conflicts on a shared file became that file’s path).rvn eval --titles 30, then --retitle for titles alone. Re-distilling the chunks instead changed contents too and could not separate the two.
Seven changes, each reviewed before it shipped, and two viewer redesigns tried and put back.
A scan of everything stored found no leaks of the formats ravnori already redacted, but 58 passwords in URLs (postgres://user:pass@host) in 45 imported memories and 13 events. Redaction now covers passwords in URLs and Stripe, Google, Hugging Face, npm, Slack-webhook and Telegram tokens. When the rules grow, the watcher re-redacts stored text once, in resumable batches.On this machine: 15 events and 61 memories rewritten in 12 s; a rescan finds none left. Transcripts on disk and older backups are left as they are.
Two time indexes, gzip and ETags, no web font, and a stylesheet cut to the rules the viewer can reach. Live updates pause in a hidden tab and catch up, oldest first, when it is shown.First load 583 KB → 50 KB, a reload 137 KB → 2 KB, a feed page on the server 370 ms → 3 ms, layout while scrolling at 4× slower CPU 710 → 233 ms. 27 feature checks pass on both builds: the rule was no capability lost.
rvn install writes each agent’s part whether the agent is installed yet or not. rvn install, rvn doctor and the viewer say which installed agents are not getting memory (missing hooks, tools pointing at a missing binary, Codex hooks not yet trusted); the viewer’s Connect button fixes it or shows the steps to trust Codex’s hooks.Re-running install leaves a working setup untouched; Codex’s config.toml is edited structurally and an invalid one is refused. Three Codex review rounds; six end-to-end tests in a scratch home.
“Back up automatically every day” can be turned off in the viewer or with rvn backup --auto off; the watcher follows it without a restart, and doctor stops alerting on an old backup while it is off. Each backup can be deleted on its own, after a confirmation.Both actions sit behind the viewer’s origin and X-Ravnori guard; delete accepts only a backup’s own name, never a path.
A crafted backup’s triggers, views and indexes are dropped over a plain connection before ravnori opens it, so none can run on ravnori’s migration writes. The filter now skips only real sqlite_* objects (a LIKE pattern had spared a trigger named sqliteevil).Found by review while building the timeline preview. The regression test plants such triggers and fails with the old order.
A lock released a moment ago could look held, because a subprocess forked by another thread keeps the lock file open until exec. Acquire now retries for 200 ms before reporting another distill.4 failures in 12 full test runs before, none in 15 after.
A three-pane viewer (filters, list, the memory in full with whether its code survived and how often agents opened it) and a live timeline at /next were both built, reviewed and run, then reverted at the user’s call: the original feed stays. The three-pane viewer is kept on branch viewer-redesign; the timeline was deleted at the user’s call, its security fix already on main.The lightweight pass above replaced them: same feed, much less weight.
One catch-up, two recall changes judged on real prompts before any code shipped, a missing alert, a live trial of how memories are shown, and what a teammate needs: distillation on their own sign-in and the service on macOS.
Sessions from 7 to 30 days ago that predated backfill were caught up by hand, project by project: 478 undistilled became 0 apart from five small or idle projects the user chose to leave as they are (103 sessions; their transcripts and events stay searchable). A first bounded run (up to 200 calls) took it to 336; then 215 sessions, project by project, took 200 calls and wrote 677 memories (478 observations, 199 summaries), all embedded. rvn doctor now reports fully caught up.A per-session dry run first: 63 of the 215 had nothing left to distill and cost nothing. Most of what remained were worker sessions with no prompt from the user, distilled at the user’s choice. A fresh install never builds this gap: the watcher backfills the 7 days before it first ran, then keeps up.
The idea: after prompt 10, show 3 memories instead of 5, require a closer meaning match, and drop or demote memories older than 7 days, since agents opened fewer of them. On the judged real prompts (70 older, 57 from the weeks ravnori distilled itself), late-session memories were judged more helpful than early ones (78–90% against 37–73%), and none of 36 variants removed more unhelpful memories than helpful ones; the best dropped 18 helpful for 2 unhelpful.Simulated on each prompt’s judged top five from rvn eval --judge --dump, before writing code. A sample of 18 sessions had suggested the opposite.
Built as a setting: when an agent opens a file, skip memories older than 7 days unless their own edited lines are still there. On the judged file edits it showed 61 helpful and 18 unhelpful memories against 68 and 22 without it, and helped 28 edits against 30: seven helpful memories lost to drop four unhelpful ones. Reverted.The same build run twice on the same edits, once with each setting, judged by the same models.
Without a distillation key, ravnori captures and searches everything but writes no memories, and nothing said so. Now that is an alert in rvn doctor, the viewer and session start, and the last line of rvn install.Tested end to end in a scratch home: said with no settings, silent with a key, and the test fails with the alert removed.
Agents use what they open, but memories shown and not opened leave no more trace in what the agent does next than memories never shown (9.0% against 9.1% at the same match strength, 528 shown over 139 prompts). Recall picks well; the presentation does not land. Since 2 October half of sessions, split by a hash of the session id, are shown at most 2 memories per prompt and the rest 5, in the same days.Result after 14 days (scripted sessions left out, opened = fetched in full within 30 minutes): shown 2, 71 sessions and 531 prompts, opened 28 of 977 memories (2.9%) and something on 3.6% of prompts; shown 5, 85 sessions and 355 prompts, opened 49 of 1,279 (3.8%) and something on 6.8% of prompts. Differences, resampling whole sessions: −1.0 points per memory (95% CI −4.6 to +2.0) and −3.2 points per prompt (−9.1 to +1.1). Showing fewer did not make agents open more, so every session is back to 5. In the same days agents fetched the full memory after 88 of 120 of their own searches (73%): what they ask for, they read.
A teammate without an endpoint or key now gets memories from their own plan: rvn install sets distillation to claude -p (sonnet), else codex exec (gpt-5.6-luna, low effort), at most 100 background requests a day, and never changes a setup that already works. Each run is isolated: an empty directory, no saved session, tools, MCP servers, hooks or user settings, thinking off. Without hooks off, Codex would run ravnori’s own hooks and every request would become a session to distil.Chosen on 30 real chunks from 19 sessions, judged blind against luna by gpt-5.6-terra and claude-haiku-4-5. Sonnet matched luna (16–13 and 13–15; 4% of specific details not in the transcript against 3%) at about $0.04 a chunk; codex with luna came close (17–13, 3%). Haiku lost 30–0 with one judge and made up more details, so it is not the default; the first suggestion had been haiku, from a single test. Thinking on cost three times more for no better memories.
One module decides the service manager: a systemd user unit on Linux and WSL, a launchd agent with KeepAlive on macOS. Install, uninstall, the health alerts and the viewer’s restart button all ask it, so each names the right command on its system.136 tests on Linux; on macOS type-checked only (cargo check and clippy for aarch64-apple-darwin). The first real run is the next step.
The first steps toward the north star: a licence that keeps ravnori open, an interface other tools can build on, and two apps on it.
Anyone may use, change and share ravnori; whoever distributes a changed ravnori, or runs one as a service, shares the change. Contributors license their work under the AGPL and to the authors, so a commercial licence stays possible. Every release now carries the licence and the texts of all 500-odd linked components’ licences, generated from what the build actually links.
A stable, read-only HTTP interface to sessions, events, memories (with the events they cite and whether their code survived) and search, so a tool reads one complete record instead of parsing transcripts. Lists page by id, so a tool can also sync. Fields are only ever added within v1. Every request carries a token from an owner-only file; a web page cannot send it, so a browser cannot read the record.On the live record (2,757 sessions, 110,630 events, 67,888 memories) lists answer in 10 ms and search in 40 ms. Tested over HTTP, including that a wrong token and a foreign Host are refused.
rvn ask: questions about past work, with sourcesThe best-matching memories are shown at once; the distillation model answers from them only and cites them by id. A citation of a memory it was not shown is dropped. Each memory carries what the developer typed in that session, so a “why” is answered from what was actually asked.Asked why the viewer redesign was reverted, the first version gave a plausible reason taken from a later next-step note; with the typed prompts and a rule against inferring motives it answers that the user asked for it, citing the right memory, in about 7 s.
rvn trace: Agent Trace for commits already madeCursor’s open Agent Trace format, built after the fact from transcripts ravnori already holds: which lines each commit added were written by an agent, by session and model. Tools that record attribution from now on cannot do this for past commits. Exact line matches only, so the share is a floor; edits in other worktrees of the same repository count.Last 40 commits, under a second each: ravnori 37% of added lines; a team repository worked through git worktrees and a merge train 14% (0% before counting worktrees and using author dates, since the merge train rebases commits hours after the edits); another 6%, where 32 of the 40 commits are other people’s.
In order of expected gain for the effort, each measured before it ships.
daefery/ravnori, and this page is on GitHub Pages. Next: the official plugin directory, and the first outside builders.rvn ask and rvn trace shipped. Next: offer the API to the authors of two or three transcript tools (standups, time sheets, cost reports) and list ravnori in Claude Code’s plugin marketplace. Shell-written files in rvn trace (heredocs, scripts) are the known gap.ravnori is the best fit for this setup: three agents on one machine, models through CLIProxyAPI, and no tolerance for silently lost sessions. Search is hybrid and measured on real prompts, including the weeks ravnori distilled itself, but Mem0, Hindsight and Graphiti still model entities and time, and agentmemory covers more agents.
The position is now built rather than planned. Capture that cannot lose data, distillation that cannot fall behind unnoticed, memories that cite the transcript events behind them and say whether the code they changed is still there, live updates across agents, and alerts that reach you. Every competitor summarises first and keeps no source; they cannot offer evidence or check an edit without storing full transcripts.
Next: toward the north star above: the record as a platform (schema, local API, ask with sources), a first run on a Mac and one teammate. Agents do not use the memories they are shown unless they open them, and showing fewer did not change that (trial, 2–16 October); what they search for, they read.
Facts marked inferred in the matrix could not be confirmed from a primary source. Prices for Mem0, Letta and Zep come from third-party listings. Roadmap order reflects council round 8 (GPT-6 Astra, Gemini Pro Agent, Claude Fable 5.1) and, for recall, rounds 12–25 (GPT-6 Astra, GPT-6 Sol, Claude Fable 5.1), and rounds 41–43 for distillation and file memories, adjudicated by Claude Opus 5.5.