Infinity Neural Memory — V1.2 (1.2.0)

A self-contained, honest, testable memory nervous system for long-running agents. Permanent memory · append-only timeline · hybrid recall · explainable governance gates · zero hard dependencies.

---

⏳ The Infinity guarantee — memory that does not evaporate

In Infinity Neural Memory, forgetting is a decision — never a side effect.

Your agent's past is permanent by design. Nothing fades because time passed. Nothing is dropped by a background cleanup. Nothing decays into oblivion.

• The timeline is append-only. What happened, stays. Events are written once and never edited.
• Contradictions supersede — they do not erase. When a new fact replaces an old one, the old truth steps aside (marked no longer current) but remains permanently on the record.
• No expiry. No TTL. No garbage collector. There is no code path that deletes a memory because it grew old, cold, or low-priority. (Verified: a full-tree search finds zero auto-deletion.)
• The only way a memory ever leaves is an explicit forget — issued by a human, for a real reason (right-to-forget, or purging a leaked secret), and itself written to the timeline as a tombstone. Erasure is deliberate, rare, and auditable.

This is what "Infinity" means: the system keeps everything you let in, for as long as you exist — until you, and only you, decide otherwise.

Permanence and sharp recall — not one at the cost of the other. Everything is kept; relevance and freshness decide what surfaces for a given query, never what survives. An old memory doesn't disappear — it waits for the right question.

For anyone analyzing this: persistence ≠ "zero-decay"

Older systems claimed "zero-decay" by freezing retrieval scores — which ranked stale memories as high as fresh ones, hurt recall, and still never guaranteed the data survived (an activation number says nothing about whether a row exists). Infinity Neural Memory keeps the two concerns separate:

Concern	Infinity Neural Memory
Persistence — does the memory survive?	Absolute. Nothing is ever auto-deleted. Only an explicit, logged forget removes anything.
Retrieval ranking — does it surface now?	Freshness + relevance decide order; pinned rules/identity never fade.

So "permanent" is a storage guarantee you can audit — not a frozen-clock trick. See docs/PERMANENCE.md for the full contract.

---

Distilled from two prior projects and two rounds of deep audit:

• infinity-neural → event-sourced timeline, instinct-recall protocol, Vietnamese NLP, dual-brain idea.
• library-memo → hybrid BM25+vector retrieval with RRF, self-contained discipline, admission concept.
• Both audits → the fixes: redaction, file-lock, real-vs-stub honesty, single-main VERIFY, provenance + supersession, memory admission gate, pre-action gate.

Design law: ship only what runs; label protocol vs code; no public claim without a passing test.

What V1 Final adds (the synthesis)

This is the merge of the self-contained engine and the OpenClaw product shell, with both audits' fixes applied:

• Two gates, not one. Write-side AdmissionGate (what to store) plus retrieval-side InjectionGate (what to inject into context). They guard different points in the lifecycle.
• Explainable Preference Prior (Brain.prefer) — the V4 "subconscious" idea as real, tested, inspectable code. No hidden bias.
• PII scrub (inmem.pii) on top of secret redaction — fixes the seed-corpus leak the lab audit found.
• Safe seed importer (scripts/seed_import.py) — bootstrap from a private corpus with PII-scrub + secret-block + dedupe; nothing raw reaches the store.
• OpenClaw protocol layer — SKILL.md, INSTINCT-PROTOCOL.md, SYNC-PROTOCOL.md, SESSION-LOGGER.md, AGENTS-BLOCK.md.
• Scale & semantic recall (V1.1). FTS5 + a two-stage pipeline (lexical candidates → vector rerank of only those → RRF) + a lazy embedding cache: warm queries ~84x faster than V1.0 and bulk insert faster than before. Swap in OpenAIEmbedder / OllamaEmbedder for semantic recall (zero new deps; the offline HashingEmbedder stays the default).
• Optional engine, honest bridge — bundles neural-memory 4.12.0 (attributed in NOTICE.md), used only via SubprocessBridge → nmem. The core never imports it; the bridge never fakes a sync.

---

What it IS

• A pure-Python (stdlib-only) library + CLI you can pip install and run anywhere — no API key, no native build, no external service required to pass tests.
• An event-sourced memory: the append-only JSONL timeline is ground truth; the SQLite store is a rebuildable projection with first-class provenance (source_event_id, valid_from/valid_to, supersedes, confidence, trust).
• Hybrid retrieval: Okapi BM25 + optional vector (offline HashingEmbedder by default), fused with Reciprocal Rank Fusion, optional reranker hook.
• Governance gates (the "fangs"):
  • AdmissionGate — explainable decision before writing (dedupe / sensitivity / confidence). Every decision returns its reasons.
  • PreActionGate — "have we failed at this before?" check before deploy/edit/config. Surfaces warnings + supporting memories; never blocks silently.
• Security by construction: a Redactor scrubs secrets on every write path; raw secrets never reach disk. The original SHA-256 is kept for dedupe without keeping the secret.
• Vietnamese-aware: tone folding + compound tokenization so quyet dinh matches quyết định.

What it is NOT

• ❌ Not a fork of neural-memory. It does not import any neural_memory API. (An optional, honest bridge can push to an external engine if you have one — see below.)
• ❌ No "zero-decay" activation trick and no ETERNAL neuron type. Permanence here is a storage guarantee (nothing is auto-deleted — see "The Infinity guarantee" above), not a frozen retrieval score. Recency/importance is handled by pinned + salience + freshness ranking.
• ❌ Not a semantic-SOTA embedder out of the box. The default HashingEmbedder is deterministic and good enough for hybrid demo/recall; swap in OpenAI/Ollama/SentenceTransformers for production quality.
• ❌ Not a Telegram bot. The digest has pluggable notifiers (ConsoleNotifier default; WebhookNotifier does a real HTTP POST). Nothing prints to console while claiming it "sent a DM".

Install & Verify

pip install -e .            # or: pip install dist/infinity_neural_memory-1.0.0-*.whl
python VERIFY.py            # 35/35 checks, exit 0 on success
pytest tests/ -q            # 64 tests

Quickstart (CLI)

python -m inmem.cli remember "Chọn Opus 4.6 vì Sonnet thiếu depth" --category decision
python -m inmem.cli log --user "Ch101 hoàn thành, score 72/80" --agent "ok"
python -m inmem.cli recall "model nào dùng viết truyện"
python -m inmem.cli preaction --action deploy --target prod      # -> high risk if a rule exists
python -m inmem.cli digest
python -m inmem.cli sync                                          # honest dry-run by default

Quickstart (API)

from inmem.brain import Brain

brain = Brain(db_path="memory.db", timeline_dir="./timeline")
brain.remember("Quy tắc: luôn redact secret trước khi log", category="instruction", pinned=True)
brain.log_exchange("Chọn FalkorDB vì graph-native", "đã ghi nhận", session_id="s1")

for hit in brain.recall("graph database"):
    print(hit.score, hit.memory["content"])

verdict = brain.pre_action("deploy", "prod")
print(verdict.risk, verdict.warnings)

Optional external bridge (honest)

The core needs no engine. If you run an external graph memory exposing an nmem CLI, SubprocessBridge will push high-salience facts to it and report real success counts. DryRunBridge (default) performs no external I/O and says so (mode="dry-run", written=0).

Known limitations (V1)

• BM25 is computed over the active set in memory — great to ~10⁴ memories; shard or move to a vector DB beyond that.
• Default embedder is a hashing trick, not semantic. Provide a real Embedder for quality recall.
• Near-duplicate similarity in the AdmissionGate is a fused-score proxy, not calibrated cosine.
• Single-process file lock (POSIX flock); cross-host concurrency needs an external lock.

Security warning

Treat the timeline + store as private by default. Redaction reduces risk but is not a guarantee — do not commit *.db or timeline/ to a public repo. Run a secret scanner in CI (see .github/workflows/ci.yml).

License

MIT — see LICENSE.