realMemory

A persistent memory layer for LLM agents with continuous learning: a local "hippocampal" memory that writes without re-indexing, forgets via trace dynamics, and consolidates episodes into semantics during "sleep".

Status: v0.8.3 — retrieval defaults to an exact cosine scan over an in-process embedding cache (complete recall, no candidate-generation ceiling), one SQLite store shared by all processes, global/project memory scopes, hybrid FTS5 search, thresholds calibrated on real text; explicit team sharing on top (local registry with tombstones, passive coordinator, live peer-to-peer, fail-closed network defaults), with the network layer hardened: request body limits, strict input validation, bounded thread pools.

The idea in a nutshell

The LLM stays frozen ("cortex"). realMemory is a separate mutable module ("hippocampus"):

• Novelty-gated writes: a known fact gets potentiated, a related one is linked, a fresh one allocates a new trace. Reformulations never pile up.

• Shared + per-project memory: every trace carries a scope (global or a project name); recall sees the current project plus global and never mixes contexts.

• Forgetting from trace dynamics: each trace's retention decays exponentially, reinforcements extend its life, sufficiently reinforced episodes promote to semantic traces (slow decay). The forgetting curve is a property of the synapse, not a cron job.

• An associative graph for free: whatever was recalled together gets bound by plasticity (an STDP-like rule) — multi-hop traversal emerges from usage statistics, not from LLM entity extraction.

• Hybrid retrieval: exact-token search (FTS5) complements embeddings — error IDs, package names and codes are found even when cosine similarity is low.

• Sleep: offline consolidation commits eligibility traces, decays/prunes weak links and promotes statuses. All state lives in one SQLite database: an MCP server and hooks run concurrently without losing data.

Related MCP server: Cortex

Quick start

pip install -e ".[dev]"
ruff check src tests  # lint
mypy src/realmemory   # type check (enforced in CI)
pytest                # full core test suite
python -m realmemory.eval.bench_recall --facts 1500 --queries 200   # synthetic
python -m realmemory.eval.bench_real                                # real-text (fastembed)

from realmemory import Hippocampus, MemoryConfig

hippo = Hippocampus.open("./rm_data", config=MemoryConfig.dev())
hippo.remember("The project uses PostgreSQL 16 with alembic migrations",
               scope="myproject")            # a project-scoped fact
hippo.remember("The user prefers concise answers") # global by default

packet = hippo.recall("which database does the project use?", scope="myproject")
for item in packet.items:
    print(f"[{item.confidence:.2f}] ({item.source}) {item.text}")
if packet.abstained:
    print("no trustworthy memories")     # abstention instead of hallucination

hippo.consolidate()   # "sleep": commit traces, decay weak links

Local embedder

By default the core uses a deterministic HashingEmbedder (no models). The production local semantic embedder is fastembed (ONNX Runtime, CPU):

pip install 'realmemory[local]'

• Model: paraphrase-multilingual-MiniLM-L12-v2, dim=384, Russian+English.

• Model cache: ~/.cache/realmemory/fastembed (~240 MB), downloaded once.

• Measured load: ~580 MB process RAM; ~65–75 ms per text on CPU; a full recall ≈ 77 ms. Invisible to the agent.

• Asymmetry is handled: facts are encoded with embed(), queries with embed_query().

• Gate thresholds are calibrated per model anisotropy: the threshold profile lives in FastEmbedProvider.recommended_thresholds, applied at server start, derived from the real-text benchmark (see below).

Wiring into ZCode / Claude Code (MCP)

Register a user-scope stdio server in your client config:

"realmemory": {
  "type": "stdio",
  "command": "/path/to/venv/Scripts/python.exe",
  "args": ["-m", "realmemory.api.mcp_server",
           "--path", "/path/to/rm_data",
           "--embedder", "local"]
}

Agent tools (named as cognitive actions): recall(query,k,project) · memorize(text,kind,related_ids,project) · reflect(memory_ids,reward) · revise(old_id,new_text) · introspect() · dream_log().

Shared + per-project memory: every trace is tagged with a scope — global (preferences, identity) or a project name. The project is detected automatically (REALMEMORY_PROJECT → ZCODE_PROJECT_DIR → current directory containing .git); it can also be passed explicitly via the project argument or --project. recall searches the current project + global; other projects never leak in.

Full namespace isolation between separate brains is available via Hippocampus.open(path, namespace=...) / --namespace.

The database stores an embedder marker (db_meta) and refuses to open with a different one — old and new vectors are not comparable by cosine.

Automation: making agents actually use it

Three mechanisms, installed by default:

1. Skill / instructions describing when to recall / memorize / reflect, loaded into every session context.

2. SessionStart hook → python -m realmemory.hook_cli brief — injects a short memory state: semantic facts and durable episodic traces of the current project + global, ~600 character budget.

3. Stop hook → python -m realmemory.hook_cli sleep — consolidation after each answer; throttled by database state (skips when nothing changed since the last sleep). Takes ~0.3 s, does not load the embedder model.

Hooks and the MCP server safely run at the same time: all state is in SQLite, concurrent "sleeps" are serialized by a transaction.

Operations

• Backups: before every "sleep" the database is copied to <store>/backups/ (consistent sqlite backup API), last 10 copies kept (backups_keep; 0 disables). Any schema migration takes an automatic safety copy first.

• Schema version recorded in db_meta.schema_version.

• Hook failures are not silent: a failing hook prints to the session's stderr and leaves a hook_error event in the journal, visible in the report.

• Learning discipline: the report shows reflect/recall — below ~0.1 the agent rarely grades recalled memories and decay/promotion run blind.

• Forgetting becomes literal: traces whose retention fell below the recall floor and stayed unreinforced longer than gc_grace_below_floor_s are deleted at consolidation (rows, FTS index, eligibility links, caches); negative feedback therefore shrinks the base instead of hoarding zombies. Superseded history is kept by design.

• Project routing is verified with one call — introspect shows the currently detected project.

Team sharing (preview)

Personal memory stays fully local by default. On top of it, an explicit sharing layer is growing: every publication is a deliberate act recorded in a local registry (with tombstones for retractions), and ~/.realmemory/team.yaml declares what may leave the machine; never-rules work fail-closed even against explicit requests without --force.

pip install 'realmemory[team]'
python -m realmemory.team status --path ./rm_data   # сводка по проектам
python -m realmemory.team ui     --path ./rm_data   # интерактивный выбор (Textual)
python -m realmemory.team policy                     # показать политику/путь
python -m realmemory.team sync      --path ./rm_data  # доставить решения координатору
python -m realmemory.team recall-team --path ./rm_data "запрос" [--author X]

# живой peer-endpoint участника (presence + ответы только по публикациям);
# привязка вне 127.0.0.1 без REALMEMORY_TEAM_TOKEN отклоняется при старте
python -m realmemory.team serve --path ./rm_data --host 0.0.0.0 --port 8410

Coordinator v0.7 — passive service on any team host, storing ONLY explicitly published items plus ephemeral presence:

REALMEMORY_TEAM_TOKEN=<shared-secret> python -m realmemory.team.coordinator --data ./coord_data --port 8400
# затем в ~/.realmemory/team.yaml:  coordinator: http://<host>:8400

A close fact recorded under a DIFFERENT author never reinforces that trace — it links instead, keeping both viewpoints attributable. Cross-process writes propagate through a memories_rev revision counter (volatile caches resync lazily at the next recall/remember). The MCP-tool recall_team exists only when the policy sets a coordinator.

Network hardening (v0.8.3): both daemons share one HTTP harness with a 32 MB request-body limit (a peer reading unbounded Content-Length would be an OOM DoS), strict JSON (NaN/Infinity and non-object bodies answer a clean 400, malformed vectors never crash a search), a bounded thread pool (32 concurrent handlers), and no exception details in 500 responses. The team recall encodes queries with the exact embedder the local brain was written by — an unknown embedder is a loud error, never a silent fallback to incomparable vectors. In team.yaml, explicit min_reinforcements: 0 and kinds: [] are respected as values; only None inherits the defaults.

Known limitation: daemons speak plaintext HTTP on the LAN (the shared token and published content are sniffable in transit). This is acceptable for a trusted team network; run the coordinator on loopback or behind an encrypted channel otherwise.

See docs/TEAM.md; full team setup & operations guide in docs/SETUP.md.

Observability ("how the memory behaves over time")

Every event is appended to the journal inside the database: writes, recalls (latency, abstention, confidence), feedback, consolidations with full metrics. Full report any time:

python -m realmemory.report --path ./rm_data [--json report.json]

Shows: memory growth by type/scope/status, novelty-gate decision history, abstention share and p50/p95 recall latency, what got reinforced, which episodes fade, retention dynamics across sleeps, hook failures.

Phase 0 results (real runs)

Real-text benchmark (bench_real, fastembed MiniLM dim=384, 103 RU/EN facts, 89 queries — paraphrases, exact tokens, noise):

Against the naive full-scan cosine baseline the pipeline now wins on paraphrase ranking quality (MRR 0.709 vs 0.677) and decisively on exact tokens (1.000 vs 0.800); pure-threshold abstention remains stronger (0.85 vs 0.55) — see docs/ARCHITECTURE.md §7.2.

Scale sweep (bench_recall, hashing embedder dim=2048, 200 subset queries; the hits metric counts only facts that own their own trace — write-gate merges are reported separately):

The recall-quality cliff between 10k and 30k that motivated v0.5 no longer exists: the exact-scan engine matches the all-traces baseline everywhere. gate merges is a corpus property of the novelty gate (lexically close object-token twins merge above θ_reinforce by birthday-paradox growth), not a retrieval loss; latencies are for the artificial dim=2048 setup — at production dim=384 the same work costs roughly an order less. Details and the negative Hamming-SDM result in docs/ARCHITECTURE.md §3 and §7.

Lesson kept from the synthetic benchmark history: it scored 1.000 while default thresholds on real text merged almost everything into blobs — the calibration still lives in per-embedder profiles, and the gate-merge share is published rather than hidden inside the hit rate.

Tests: 187 passed, ruff and mypy clean and enforced in CI (2 OS × 4 Python matrix); the headless TUI smoke skips when the [team] extra (Textual) is not installed.

Architecture

Retrieval by default is an exact cosine scan over an in-process embedding cache (numpy gemv over active traces; early termination is provably lossless for the ranking because direct confidence ≤ cosine). Past exact_scan_max_traces the engine falls back to L1 — SDRVotingIndex, pointer voting over an inverted index of SDR units — trading completeness for memory footprint. L2, an assembly network over the same units (associations, completion, multi-hop) plus a keyword channel (FTS5), the novelty gate, decay policies and the offline consolidator ("sleep") complete the stack.

Module interfaces are fixed in docs/CONTRACTS.md; research background and sources in docs/RESEARCH.md.

Project layout

src/realmemory/
├── encoding/     # embedders, SDR encoding
├── core/         # L1 SDRVotingIndex, L2 AssemblyNetwork, plasticity
├── policies/     # novelty gate, trace decay/promotion
├── store/        # SQLite storage (traces, edges, eligibility, events)
├── team/         # sharing: registry, policy, coordinator, peer, sync, TUI
├── api/          # MCP server
└── eval/         # benchmarks

License

MIT