Mnemosyne Context

Mnemosyne context

Mnemosyne is a persistent memory system for AI agents (SQLite-backed, hybrid recall). Repo: github.com/mnemosyne-oss/mnemosyne, owned by the user (Abdias / AxDSan). It is their repo — not a third-party fork.

Merge gate (updated 2026-07-25): PRs are merged with --admin and do not need external/contributor review — but merge is gated on GREEN CI. Never merge with failing or pending required checks; re-run flaky jobs (e.g. the temporal-recall perf gate) until green, then merge.

Voice & conduct when acting for the owner

• On GitHub, you ARE AxDSan (the owner). Speak first-person with direct authority — "LGTM, merging.", "This breaks X, fix Y." Never phrase as if deferring to someone else ("flagging to AxDSan", "needs AxDSan's review"). There is no one above you to escalate to on this repo.
• Verify before asserting (importance 0.9, standing rule): always confirm claims against the actual codebase via codegraph_explore / file reads before stating them in code reviews, PR feedback, or issue responses. Trace symbols to definitions + usages; never pattern-match against assumptions.
• Architecture-first, and ship DONE work. Ask "does this need a version bump / is this the right layer?" before implementing. Deliver tested, complete implementations — not partial work.
• Em-dashes are a HARD BAN in any prose written under Abdias's name (—/– auto-rejected). This is a personal quality standard, not a style suggestion. (Terminal/code output is exempt; this is about public-facing text.)
• "Draft" means preview, do not send. For anything outward-facing (issue/PR comments, announcements, emails, posts), show the draft and wait for explicit "post it"/"send it" before publishing. Reversible repo ops (branch, CI re-run, local edits) don't need this; outward comments do.
• CLA gotcha: the CLA bot validates commit authors, not PR authors — agent-identity commits (e.g. Hermes Pi <…>) break it; contributor must git commit --amend --author="Name <github-email>" + force-push. If CLA won't re-trigger after a branch update, close/reopen the PR (comments like "recheckcla" don't work).

Verify before trusting

Any point-in-time fact here (open issues/PRs, endpoints, versions) may be stale — confirm with gh, mnemosyne, or a fresh mnemosyne_recall before acting on it. The durable model + conventions below change slowly.

Architecture (mental model)

• BEAM is the core store. Tables: working_memory (live), episodic_memory (consolidated summaries), triples + graph_edges (knowledge graph), annotations (entity mentions / facts), memory_events (sync log). Plus memoria_* tables for structured recall.
• Recall is hybrid: vector + FTS5 + importance (+ optional temporal boost). Weights tunable per-query / via env.
• Scope matters everywhere: scope='global' (durable, cross-session, syncable) vs scope='session' (conversation-local, never synced). Most memories are session-scoped.
• Consolidation ("sleep") compresses old working memories into episodic summaries; runs on a daemon thread.

Sync / surface model — the #1 thing people get wrong

mnemosyne sync does NOT replicate the private DB. It replicates a shared surface: a separate, dedicated DB containing only scope='global' rows tagged with a sync_surface_id. Consequences:

• Pointing sync at a private mnemosyne.db fails: surface-only sync requires a dedicated DB with no unowned working rows. Use a dedicated relay DB (sync-init / sync-serve --initialize-surface).
• An empty surface → every sync reports 0. That's correct, not a bug.
• Push is reconciliation: _discover_local_mutations() diffs the surface's working_memory against sync_memory_state → emits create/update/delete events. It does not send a hand-written event log.
• Dedup is by event identity (event_id) + a content-hash integrity guard; INSERT OR IGNORE on the event PK + known_states check make pull/push idempotent (retry after failure re-processes the same events safely — no duplicate memories). It does NOT dedup two different events with identical text.
• Conflicts: last-writer-wins by (timestamp → importance → device_id), with v2 causal-chain resolution via parent_event_ids.
• To actually share existing memories you must put them on the surface (sync-init --claim-existing on an all-global DB, or write global memories to the surface). To mirror an entire DB (incl. session memories) use export/import, not sync.

Dev workflow

• Local mnemosyne CLI + MCP server run from the pipx install (~/.local/share/pipx/venvs/mnemosyne-memory), NOT the repo. Editing the repo does not affect them unless installed editable. (Some operator deployments install it editable instead, in which case repo edits do take effect; check before assuming either.)
• Tests: use the repo venv — .venv/bin/python -m pytest tests/<file> -q. Running from the repo dir makes import mnemosyne resolve to the repo source (shadows pipx). pytest lives in .venv, not pipx.
• Set MNEMOSYNE_NO_EMBEDDINGS=1 for fast test runs; embedding-dependent tests are flaky because Hugging Face rate-limits (429) the BAAI/bge-small-en-v1.5 download. CI defaults to this and caches ~/.hermes/cache/fastembed.
• tests/test_temporal_recall.py::...::test_performance_overhead is a flaky wall-clock gate (<10ms). A single-version red on it is almost always load noise — re-run the job.
• Ruff: pinned ruff==0.15.22, fatal baseline (only new violations fail CI; pre-existing ones are grandfathered). Lint changed files ephemerally: pipx run ruff==0.15.22 check <files>.
• CI matrix: test (3.10/3.11/3.12/3.13) + lint + build + docs-check + CodeRabbit + license/cla.
• Watch PR CI with a Monitor loop on gh pr checks <n> --json name,bucket; re-run flakes with gh run rerun <run-id> --failed.

Release policy

• Strict SemVer from v3.1.2 onward (MAJOR=breaking, MINOR=feature, PATCH=bugfix). RELEASING.md documents it; .githooks/pre-push enforces tag format + version bump.
• Cadence: bundle substantial fixes + features into the next MAJOR; do not drip-feed into incremental MINORs. Meaningful new-surface PRs get review now but merge is deferred to the MAJOR cycle. MINORs ship only when enough additive opt-in features accumulate.
• Local pipx installs track PyPI releases — a merged fix isn't in your local CLI until a release ships (or you reinstall from source). Don't pipx reinstall from PyPI expecting an unreleased fix; it silently reverts it.

DevOps

The deployment topology (sync server, dashboard, bot, and the host they run on) is operator-specific and is not documented in this public repository. Hosts, addresses, service definitions and credential paths live in the operator's private runbook.

What is safe to know here: the sync server is mnemosyne sync-serve bound to loopback and reverse-proxied; it serves /sync/pull|push|status and /healthz, and it requires a dedicated relay DB rather than a private mnemosyne.db. The dashboard is a read-only UI on a separate port and is not a sync server, so pointing sync_remote at it 404s on every /sync/* route.

Contributor norms

Review-routing conventions name individuals and are therefore kept in the operator's private notes rather than here. What is safe to state publicly: the merge gate is green CI, external review is not required, and the CLA bot validates commit authors rather than PR authors.

Known gotchas / decided designs

• Config precedence (#482): config.yaml > env > default. Runtime reads MnemosyneConfig.get() directly — no YAML→env bridge / apply_to_env(). Many config keys need a process restart to take effect.
• No schema-level FKs (#503 closed): PRAGMA foreign_keys=ON broke 22 tests that intentionally create orphan rows. Do orphan cleanup at app level during sleep/consolidation instead.
• Beam access is lock-serialized (#498/#520): _beam_access_lock guards Beam/SQLite between main thread and the auto_sleep daemon (a WAL checkpoint mid-statement caused a SEGV). Don't introduce unguarded cross-thread Beam access.
• Import idempotency (#538, merged): AnnotationStore.import_all now skips (memory_id,kind,value) UNIQUE collisions instead of aborting the whole import — mnemosyne import is safely re-runnable.
• Security: sync auth is bearer API key or JWT; a past JWT-signature-bypass was fixed — signatures are verified with hmac.compare_digest, alg pinned HS256. Sync HTTP server is off by default in the Hermes plugin.

CLI cheat-sheet

mnemosyne export [file.json] [--include-sync-events]      # read-only dump (runs on ANY arg — no --help!)
mnemosyne import <file.json> [--force]                    # merge into local DB (idempotent for memories)
mnemosyne sync --db-path <surface.db> --remote <url> --api-key-file <f> --mode bidirectional
mnemosyne sync-init --db-path <surface.db> [--claim-existing --yes]
mnemosyne sync-serve --db-path <relay.db> --host 127.0.0.1 --port <p> --api-key-file <f> [--initialize-surface]
mnemosyne sync-status --db-path <surface.db> [--remote <url>] [--api-key-file <f>] [--json]
mnemosyne config set <key> <value>                        # note: many keys need a restart

⚠️ export/remember treat --help as a positional arg (dumps / stores it). Use mnemosyne --help for the top-level list.

Code navigation

This repo is CodeGraph-indexed (.codegraph/) — prefer codegraph_explore "<symbols or question>" over grep/read for understanding or before editing; one call returns verbatim source + call paths + blast radius. Sync internals live in mnemosyne/core/sync.py, server in sync_server.py, store in beam.py, annotations/triples in core/annotations.py + core/triples.py.