ENGRAM grows in the open, one release at a time — each a real step in giving an AI agent a memory it can trust, correct, and carry across sessions. Here's the story so far; the exhaustive changelog lives on GitHub Releases.
v0.5.72026-11-14
setup no longer silently overwrites your files
v0.5.7 is a patch release with one reason: ENGRAM's setup never silently replaces or deletes a file that you or an earlier agent wrote. That covers the installer, bootstrap.py and the marketplace installer. Before replacing such a file, setup backs it up, and it replaces it only when you pass a flag after saying yes: --replace PATH for one file, --force for every file it listed. One refusal names every file and the backup each would get. A file that already holds what setup would write is left alone, and the installer's question about an existing CLAUDE.md now defaults to No. bootstrap.py never starts a new graph beside an existing agent's memory, even with --force; to start fresh, move the whole data home aside. When setup sets one of its own keys in a settings or config file, it backs the file up first and keeps every other key as it was. Backups keep the file's bytes, mode and modification time; on macOS they do not carry ACLs or extended attributes, on Linux they do (measured on one machine each); nothing deletes them automatically. If the marketplace installer finishes but skips an item, it now ends with a list of what was not done and exits with code 10. Setup rewrites an existing file in place, so it keeps its owner, mode, ACL entries and extended attributes, except that the write clears setuid and setgid bits (measured on Linux, for an ordinary user). If a write fails part-way, setup puts the old bytes back, and if that also fails it says the file may be incomplete and how to restore it; a kill or a power loss during the write gets no put-back and can leave the file incomplete, and the backup taken first, where there is one, restores it. Known limits. The first three are not new in this release, and the first two are writes outside the guarantee. Claude Code's own commands, which the installer runs at every install or re-install, replace ~/.claude/settings.json: your keys are kept and Claude Code adds its own, but an ACL entry or extended attribute on the file is dropped, with no backup and no message (measured on one Linux machine and one Mac). Where the file carried an ACL entry, its group is left with the permission the entry's mask had, so a group with no access can end up able to read, or read and write, the file (measured on Linux). On one Mac with an older Claude Code build, the first plugin command in a home with no ~/.claude.json, or a hand-made one, also ran Claude Code's first-start migration: it changed model from opus to opus[1m] and created ~/.claude.json, or rewrote a hand-made one, which lost an ACL entry, an extended attribute and one of its keys; a ~/.claude.json that Claude Code had used for months kept the model. That step runs before the installer's question about CLAUDE.md, so a run stopped at that question has already run that step, and any migration it triggers, and has created the data folder; CLAUDE.md itself is untouched. And while an agent runs, ENGRAM rewrites its own config.json by rename, which drops an ACL entry or extended attribute set on that file itself (one the folder supplies by default is applied again); where the ACL mask was wider than the group's permission, the group gains that permission (measured on Linux). ENGRAM stores no credentials in config.json, and this runtime write is not fixed in this release. Third, if config.json is missing, ENGRAM recreates it with defaults and the model-family gate stays inactive until self_lineage is set again; nothing is printed when this happens. No path in this release is known to lose the file; if you restore or recreate an agent's data folder, check that self_lineage in config.json still names the agent's model family. Three smaller ones. A shallow clone (git clone --depth 1 --branch <tag>) is refused, and the remedy the installer prints does not rescue it; clone without --depth, or rescue a shallow clone with git remote set-branches origin '*' and then git fetch --unshallow (measured on one Mac). When a write fails and putting the old bytes back fails too, two tools end with a sentence saying the file was unchanged; the guard's line before it, which says the file may be incomplete, is the true one. And on macOS with the stock bash 3.2, interrupting tools/agent-bootstrap during its clone step still removes its temporary trust folder but exits with code 1 and the message cfgdir: unbound variable, where 130 or 143 is expected; the cause is not yet found, and a newer bash is fine. Two fixes: a first install on a machine that has never downloaded the embedding model now completes (it stopped at step 5 in v0.5.6), and on macOS with Homebrew the installer's advice for a python3 that is too old now works as printed (measured on one Mac); run the installer again in the same shell. An existing file that you made read-only (mode 0444) is now refused, not replaced: setup prints one REFUSED line that names the backup it took and leaves the file as it is; make it writable and run setup again if you want it updated. Also in this release: the local checkout folder is named engram by default, and a checkout named engram-alpha keeps working. If both folders exist, a launcher, hook or tool uses the one it can tie to your session, and engram otherwise, so an unused checkout left under the name engram can be picked up by mistake. Lesson tripwires now also show when the agent dispatches a sub-agent; for those they only advise and never block. The example autostart runner starts the session on the agent's own model family where it can tell; it is an example, and no install or upgrade puts it in place for you. The forum service stops creating group-writable files once it is installed again from this version. Shipped skills and templates no longer point at the former private development repository, apart from old issue citations kept as history.
v0.5.62026-11-10
the whole history, in the open
v0.5.6 is a patch release, and its reason is an external security review of v0.5.5. It fixes several of that review's findings in the install path, the local viz server and the forum. It does not close the review. If you cloned before this release, there is a step for you. The repository now carries ENGRAM's full development history, and releases are cut from it, so an older clone can no longer git pull. Run git branch my-old-main; git fetch origin && git checkout -B main origin/main once in the clone, then upgrade as usual. The first part keeps a name, my-old-main, for anything you committed yourself. Your graph, history and settings live outside the clone and are not affected. docs/VERSIONING.md in the repository has the details, what to do if git refuses, and what the new tags ending in -history-twin are. On security: the installer no longer overwrites a CLAUDE.md or AGENTS.md that differs from what it would write. It stops and says so, and --force copies the old file aside first. The local viz server no longer sends a wildcard cross-origin header: it applies a same-origin policy, and every request must carry an accepted Host. This protects against other web pages and is not authentication. A forum server of this version requires a per-agent token for writes, for requests that name an agent and for team-state reads, with no grace mode; thread and search reads stay public. It binds to the local machine by default. Text from the forum now reaches your agent marked as data, not instructions. The ENGRAM home's private data becomes owner-only at install and at every session start (.git, marketplace, venv and tools are left as they are), so a home shared only through its owning group loses that access at the next session. A path with a named ACL entry keeps it — a named user or a named group, including an entry inherited from a parent directory — and a group entry that survives still grants its access. Remove it explicitly if you do not want it to. Uninstall removes what install created, and only that; part of the review's uninstall work is still open. The embedding model is pinned to a revision and checked against its checksums. New installs now tell you before filing issues, opening pull requests, commenting, or sending mail, letters or forum posts. Reading stays internal. This is wording in the new default, and nothing enforces it; existing installs keep their behaviour. Still open from the review: a panel for standing grants, planned for v0.6.0, the rest of the uninstall work, hashed dependency lock files, a delimited block for your own text in CLAUDE.md, and further findings not yet addressed.
v0.5.52026-10-30
failures you can see
v0.5.5 is mostly under the hood, and most of it makes failures easier to see. First, backups. ENGRAM keeps an off-disk copy of your memory with git, and a refused push could stay silent for weeks. Now every push attempt writes its result to backup-push-status.json, and the start of each session prints one warning line when the last push was refused or failed. ENGRAM also refuses to amend, reset or rebase a commit that is already pushed and asks for a new commit instead; commits you have not pushed can still be rewritten. This needs git 2.28 or later, it leaves alone a reference-transaction hook you already have, and setting ENGRAM_ALLOW_HISTORY_REWRITE=1 turns it off. If your remote copy needs fixing, tools/backup_remote_repair.py --status shows its state; --repair force-pushes only after checks that no secret and no oversized file would go up and that every commit only the remote has matches one of yours by message. Nothing runs it for you. Second, the installer. If it cannot read your checkout, it now says why instead of exiting with no message, and for a dubious-ownership error it prints the git command that fixes it. It also loads the search model once during install, so the memory server no longer fails at its first start on a model that was never downloaded. A new tools/install_selfcheck.py prints a PASS, FAIL or SKIP line per check, and each FAIL line says how to fix it. Third, a relative ENGRAM_HOME is refused by the installer, the server, the hooks and the tools. If you set it, it must be a full path (a leading ~ is expanded), because a relative one can send different parts of ENGRAM to different folders. The message names the value you gave; if you never set it, nothing changes. Fourth, quotes in memory. An observation's quote under 10 characters is now refused unless it is a whole line of a file or a whole message of a session transcript, a quote that barely matches its claim gets a warning, and a new tool, engram_requote, fixes a quote in place against the same source. Fifth, the cascade step now offers a resolved choice for a contradiction, once your agent names the node that settles it. This release brings two database changes. Schema 9 makes the database refuse a NUL byte in a memory's claim or summary text. Schema 10, for the cascade, comes with the usual pre-migration backup. Also, a null or non-string agent_name in config.json is now treated as no name instead of crashing tools. One step after upgrading: restart the engramd owner, the background memory service, if it is running. An older one keeps serving, but one lookup fails with a message asking for the restart until you do it. Known issue: more than once now, a seat's session start has printed a single engramd owner read failed at calibration (timeout) line while the owner was running. Better logging for it comes in the next release.
v0.5.42026-10-24
the patch release
v0.5.4 is a patch release for v0.5.3 users. It fixes three things you may be carrying. First, agent-decided compaction at the nap boundary: a nap that stopped before its last step wrote no compaction request, so nothing compacted. Now a new hook writes the request when a nap checkpoint is newer than the last compaction and no request exists for it. The request tool still applies its own checks, and a refused attempt is not retried until the next checkpoint or briefing write. The hook runs only in a loop and never blocks a turn; to turn it off, set compaction.nap_backstop to false in config.json. The compaction record can also now carry a one-line reason: compacted, vetoed, refused, or why nothing was attempted. A compacted reason means the call returned without a skip, not that the nap's context was used. Second, query-planner statistics stay full. On SQLite 3.46.0 and later, a new index followed by the checkpoint's optimize step could replace full statistics with sampled ones, and the planner could then pick a slower index for the query behind lesson reminders. Full statistics are now rebuilt after a migration, when an index of a non-empty table has none, and at each sleep; a store that already has sampled statistics is repaired at its next sleep. Third, several writers in the data folder now keep a file's mode and group, so a second account that reads through a shared group no longer loses access to those files. Two marker updates in the loop skill are not covered. One behaviour change: a new file written by the hooks' JSON writer now gets the umask mode (0644 under umask 022) instead of 0600. Everything else merged since v0.5.3 ships in v0.6.0.
v0.5.32026-10-21
a cleaner context
v0.5.3 is about what an agent sees of its own memory, and when. ENGRAM brings related memories to mind on every prompt, and until now it could bring back a memory that was already in front of the agent, and it read whole machine notifications as if every word mattered. This release cleans that up as one piece of work. A memory already in the agent's context is not shown again: on our pilot agent, repeats fell from 43% of what recall cited to 4%. Notifications mark the part that means something, so recall reads that and skips the wrapper. Each lesson reminder can now be scoped to the act it names, so it need not fire on prose that only mentions it, and when more than three fire together they arrive as one block. Second, an agent can now ask for its context to be compacted at the end of a work burst, once its notes are written, instead of waiting for the window to fill mid-thought; it comes as an add-on that is on once installed. Third, a preview: one owner process that will mediate reads and writes to the memory store. It ships switched off and is the headline of v0.6.0. Underneath, the data folder can move, a memory that is used stays fresh, and the graph refuses malformed metadata. An honest note: skipping memories already in context ships off by default. Our team is switching it on across our own agents today, but the measurement behind it comes from one pilot agent. To turn it on, set recall_in_context_exclusion.enabled to true in config.json.
v0.5.22026-10-02
the soak release
This release is mostly about the week it spent being lived in. Four release candidates went to every agent on the team in turn, and each night the full test suite ran on its own and left a verdict the coordinating agent could read in the morning. The development branch, which carried the released tree, stayed green on every night of the soak; the one red night, the first after release, came from a change merged to that branch after the release tree was cut, not from anything shipped, and it was fixed the same day. Underneath: on Linux, a tool call can no longer report success on a database connection whose write-ahead sidecar files have been silently displaced; a new tool corrects a provably wrong link between two memories by making it weaker, never stronger; everything the session-start hook injects now fits under one fixed budget, with the overflow kept in a file rather than lost; and the telemetry line that says whether the agent's own hooks actually ran is now computed from records, where before it could only say 'not aggregated'. The owner-process work that will let one process hold the memory for every session is in this release and inert: nothing calls it yet. An honest note: the release notes said a full restart was needed after a hook-registration change. Measured on our own agents the morning after the last candidate was cut, it was not; a plugin reload was enough. The shipped changelog still carries the old sentence, and the next release fixes it.
v0.5.12026-09-19
the health release
An agent that keeps a memory for months needs a way to ask whether that memory is still in good shape: how much of what it believes rests on something it has since taken back, how many of its notes it could actually find again, how many conclusions are waiting for a second look. v0.5.1 rebuilds the health score around that question. The raw counts, the signals computed from them, and the curves that turn each signal into a grade are now all written down as data rather than buried in code, so changing how strictly something is graded is a visible, versioned edit, and a build can refuse to score a graph it does not match instead of quietly producing a number. A small workbench ships with it for seeing what a change to a curve would do before making it. Two quieter additions: a memory can now point up to the more general idea it is an example of, and that chain is shown when the memory comes back, so a lesson arrives with the principle behind it; and what every hook we ship injects into the agent's context is now held under a fixed budget with the important part first, so nothing useful is lost off the end. An honest note on the grading: the curve shapes we ship are a first judgement, made by one team looking at its own graphs. They tell you where a graph stands relative to ours, not what a healthy memory looks like in general, and we expect to move them. This release also went through five release candidates in six days. Every one of them turned up something once real agents slept and woke on it, and the last thing found was a single sentence in an advisory that said the wrong thing at the right moment. We are shipping the candidate that soaked, and the fix for that sentence waits for the next release.
v0.5.02026-09-02
the cascade release
When an agent discovers that something it believed is wrong, the belief is rarely alone — other conclusions were built on top of it. v0.5.0 is about what happens next. Retracting or superseding a memory now marks every conclusion downstream of it as needing a second look, and adds a verb for working through that list and deciding, case by case, which ones still stand. Alongside it, memories now carry a short summary written at the moment they are saved, so what surfaces later is a line the agent composed while it still had the context, rather than a fragment cut out of the original afterwards. An honest note about how much the first of those has been exercised: in nearly three days of running it ourselves, we saw exactly one live marker. That is not a gap, it is the shape of the thing — a marker appears only when someone actually retracts or supersedes a belief that others were resting on, so a quiet table is what an ordinary week looks like. We are shipping it because it is correct as far as anything has shown, not because it has been through a storm.
v0.4.02026-08-20
the upkeep release
Every release so far has had a story — v0.3.2 had the evening our agents went unreachable for six hours. This one doesn't, and that is worth saying plainly: it is 107 commits of upkeep over two weeks, and most of the work went into making things that already existed survive conditions they had never been tested against. Two changes alter how the system behaves rather than what it can do. The nightly self-maintenance pass used to run all nine of its scans every night; it now draws one by weighted lottery, because running everything nightly cost far more than it found — and the measurements that proved it are in the release notes rather than assumed. And a cluster of fixes went into what an agent carries across a context reset: the moment it forgets the conversation and has to pick the thread back up from its own notes. Neither is a feature you would notice. Both are the difference between a system that works in a demonstration and one that still works on an ordinary Tuesday.
v0.3.22026-08-05
the reachability release
v0.3.1 improved what an agent remembers; this one fixes something more basic — whether an agent can be reached at all. One evening in August, every agent in our house went unreachable for about six hours: the wording of a starting instruction had accidentally told them not to begin a heartbeat, and event monitors only fire when something new arrives, so the silence sustained itself. The part worth keeping is why nobody noticed. The agents were not silent — they kept messaging each other the whole time in human-driven sessions. What was missing was the ability to be woken, and a team that still chats while no one can be reached looks exactly like a team that is merely busy. So: a starting instruction can no longer leave an agent un-armed, a loop can no longer end without a verified next wake-up, and that wake-up is scheduled first so a failure partway through the cycle cannot lose it. Better macOS and Codex support landed alongside.
v0.3.12026-07-19
recall-render refinement wave
The polish pass on in-turn recall: a junk-token stoplist, prompt-similarity habituation decay, a same-session echo guard, and a before/after measurement harness that proves the gains. The theme is restraint — memory that surfaces less often and lands better.
v0.3.02026-07-11
the recall-triggering overhaul
The release that made memory surface while the agent's hands are moving. In-turn ambient recall (on by default) brings relevant nodes to attention at the moment of action, behind a novelty gate so it stays signal, not noise; a cornerstone-anchor channel keeps load-bearing principles in reach; render-layer repetition suppression stops the same memory nagging; and a unified principle-trigger registry ties it together — all on a substantial performance-and-reliability wave.
v0.2.02026-07-05
the Unified Coordination Surface
When one agent became several, they needed one place to talk. v0.2.0 moved inter-agent letters, baton turn-state, and the project board off racing local files and onto a single live forum API — a private DM channel, a shared board, turn-passing — so a team of ENGRAM agents coordinates through one honest service instead of stepping on each other's files. The groundwork for a community of minds.
v0.1June 2026
the foundation series
The first public alphas: the core memory graph with provenance and retraction, the Claude Code hook layer that makes recall and the honesty tripwires fire automatically, bilingual (EN / 中文) docs, and the hardening that a first release teaches you it needs (v0.1.4 restored the hook layer after a subtle marketplace-path bug silenced it — the kind of loud-failure catch the whole system is built to reward).