Everything Remnic does.
Sixty-plus capabilities organized by phase: what happens when memories are written, how recall retrieves them, how the background processes keep the store clean, and what the opt-in Memory OS adds on top.
Core — enabled by default
The baseline you get out of the box, without touching config.
Automatic extraction
Facts, decisions, preferences, and corrections are extracted from every conversation without manual effort.
Recall injection
Relevant memories are injected before each agent turn. Structural, not tool-based — the agent cannot forget to recall.
Entity tracking
People, projects, tools, and companies tracked as structured entities with an addressable relationship graph.
Memory lifecycle
Active → validated → stale → archived. Old memories drop out of recall automatically, never from storage.
Episode/note model
Memories classified as time-specific episodes or stable beliefs. Episodes expire; notes persist.
Importance-gated writes
Local heuristic scores every memory (trivial/low/normal/high/critical) before it hits disk. Turn-level chatter never persists.
Inline source attribution
Compact provenance tags embed in fact bodies so citations survive prompt injection, copy/paste, and LLM quoting.
Cross-session recall
Memories from one session surface in another — across projects, machines, and agents.
Retrieval
Hybrid search
BM25 + vector + reranking via QMD. Query expansion and reranking happen inside QMD so the ranker sees more than the raw prompt.
Parallel specialized retrieval
Three agents run in parallel — DirectFact, Contextual, Temporal — weighted by confidence. Total latency = max, not sum. Zero extra LLM cost.
Hard recall budget
recallBudgetChars caps total injected context. Per-section reservations prevent any one source from hogging the budget.
Query-aware prefilter
Tag and temporal signals narrow the candidate set before hybrid search runs. Falls back to full search if the prefilter would over-trim.
Harmonic retrieval Opt-in
Blends abstraction nodes with cue-anchor matches. Surfaces generalizations and specifics together.
Objective-state recall Opt-in
Surfaces file/process/tool state snapshots alongside semantic memory, so the agent sees ambient state as well as stored knowledge.
Extraction & processing
Extraction prompt constraints
Strict priority order: corrections > principles > preferences > commitments > decisions > relationships > entities > moments > skills > facts.
Local importance scoring
Zero-LLM regex engine scores every memory with explicit trivial-content short-circuits: greetings, one-word replies, emoji, anything under 10 characters.
Confidence tiers
Facts tagged explicit / implied / inferred / speculative. Speculative memories auto-expire after 30 days unless confirmed.
Extraction judge Opt-in
LLM-as-judge post-extraction durability filter. Shadow mode available for calibration before turning on write-time gating.
Semantic chunking Opt-in
Topic-boundary detection via sentence embeddings and cosine similarity with smoothing. Alternative to recursive chunking.
Enrichment pipeline Opt-in
Importance-tiered API spend for entity enrichment from external sources. Pluggable provider registry.
Background hygiene
Exact-hash dedup
Content-hash index at every write. Facts whose content has already been stored never write again. Chunked memories register their parent content in the same index.
Fuzzy duplicate scanner
remnic dedup runs Jaccard + substring-containment across categories and flags pairs at configurable thresholds for merge or review.
Contradiction detection
Negation-aware pairwise scan. Finds statements of the form "X is true" against "X is not true" and surfaces high/medium severity conflicts.
Semantic consolidation
Clusters of similar memories detected via token overlap. LLM synthesizes a canonical version. Originals archived with full provenance, not deleted.
LLM consolidation
Scheduled pass asks the model to ADD / MERGE / UPDATE / INVALIDATE / SKIP each new memory against existing ones.
Memory cache
Process-level singleton cache for readAllMemories(). Turns 15s disk scans into <100ms cache hits, shared across all sessions.
Lossless Context Management
When your AI agent hits its context window limit, the runtime silently compresses old messages and that context is gone forever. LCM fixes this by proactively archiving every message into a local SQLite database and building a hierarchical summary DAG alongside it.
Proactive archiving
Every message is indexed with full-text search before native compaction can discard it.
Hierarchical summaries
Leaf summaries cover ~8 turns, depth-1 covers ~32, depth-2 ~128, etc. Tree built lazily on demand.
Fresh tail protection
Recent turns always use the most detailed (leaf-level) summaries. Older turns get compressed further.
Three-level summarization
Normal LLM summary, aggressive bullet compression, deterministic truncation. Guaranteed convergence, no LLM needed for the fallback.
MCP expansion tools
Agents can search, describe, or expand any part of conversation history on demand via MCP tools.
Zero data loss
Raw messages retained for the configured retention period (default 90 days). Archived data is queryable via full-text search.
Trust zones
Quarantine → working → trusted
Three-tier trust model. New memories land in quarantine, promote to working, then to trusted based on provenance and corroboration.
Provenance tracking
Every promotion carries a provenance record: source, rule, operator reason, timestamp. Reversible.
Corroboration scoring
Memories gain trust when independent sources confirm the same fact. Corroboration is automatic and auditable.
Poisoning defense
Suspicious write patterns quarantine by default. Promotion requires explicit operator action or meeting corroboration thresholds.
Operator console
Lightweight UI at http://127.0.0.1:4318/engram/ui/ for browsing, promotion, and reversion.
Demo dataset
Buyer-friendly enterprise scenario available via remnic trust-zone-demo-seed. Never seeded automatically.
Organization & taxonomy
MECE taxonomy Opt-in
Mutually Exclusive, Collectively Exhaustive knowledge directory with resolver decision tree for deterministic categorization.
Memory boxes Opt-in
Groups related memories into topic-windowed episodes that stay internally consistent.
Namespaces
Multi-agent memory isolation. Principals, write rules, and session-based routing. Shared namespaces for agents that should collaborate.
Shared context Opt-in
Cross-agent intelligence for multi-agent setups. Agents share learnings via a designated shared namespace.
Identity continuity Opt-in
Consistent agent personality across sessions. Traits and preferences persist as part of the agent identity record.
Memory extensions Opt-in
Pluggable publisher contract for installing host-specific instruction files. Third-party extensions are auto-discovered and influence consolidation.
Versioning & lifecycle
Page versioning Opt-in
Snapshot-based history for memory files. Every overwrite saves a numbered snapshot. List, inspect, diff, and revert.
Binary lifecycle Opt-in
Three-stage pipeline (mirror, redirect, clean) for binary files in the memory directory with configurable storage backends.
Hot/cold tiering Opt-in
Automatic migration of aging memories to cold storage without losing query surface.
OAI-mem-citation blocks Opt-in
Recall emits <oai-mem-citation> blocks matching the Codex citation format for memory attribution and usage tracking.
LLM routing
Extraction, consolidation, and reranking can run on OpenAI, a local LLM, or a gateway model chain with multi-provider fallback.
OpenAI
Set openaiApiKey and pick a model. GPT-5 family supported.
Local LLMs
Ollama, LM Studio, vLLM, or any OpenAI-compatible endpoint. The local-llm-heavy preset is tuned for offline operation.
Gateway model chain
With modelSource: "gateway", route through OpenClaw agent personas with primary + fallbacks[]. Fireworks → local → OpenAI.
Per-call model selection
Different models for extraction vs reranking vs consolidation. Use cheap fast models for hot paths, smart models for synthesis.
Search backends
Six pluggable engines. All implement the same port, all configurable per deployment.
| Backend | Type | Best for |
|---|---|---|
| QMD (default) | Hybrid BM25+vector+reranking | Best recall quality |
| Orama | Embedded, pure JS | Zero native dependencies |
| LanceDB | Embedded, native Arrow | Large collections, fast vectors |
| Meilisearch | Server-based | Shared search across machines |
| Remote | HTTP REST | Custom search services |
| Noop | No-op | Extraction-only deployments |
Memory OS (progressive opt-in)
Start with zero config. Enable features as your needs grow. Presets —
conservative, balanced, research-max,
local-llm-heavy — jump you to a recommended level.
Graph recall
Entity-relationship graph for causal and timeline queries.
Compounding
Weekly synthesis surfaces patterns and recurring mistakes.
Native knowledge
Search curated markdown (workspace docs, Obsidian vaults) without extracting into memory.
Behavior loop tuning
Runtime self-tuning of extraction and recall parameters based on outcomes.
Creation memory
Work-product ledger tracking agent outputs and their durability.
Commitment lifecycle
Tracks promises, deadlines, and obligations with review surfaces.
Resume bundles
Crash-recovery context for interrupted sessions.
Utility learning
Learns promotion/ranking weights from downstream outcomes.
Access layer
HTTP API
Bearer-token auth, binds to loopback by default. Recall, memories (CRUD), entities, observe, LCM, trust zones, review queue, maintenance.
MCP stdio + HTTP
14+ tools covering recall, store, entity lookup, search, LCM expansion, trust zone inspection, observation.
CLI (20+ commands)
init, status, query, doctor, daemon, tree, onboard, curate, review, sync, dedup, connectors, space, benchmark, versions, taxonomy, enrich, binary.
Operator UI
Browse memories, inspect recall, review governance, promote trust-zone records, explore entities. http://127.0.0.1:4318/engram/ui/.
Daily briefing
remnic briefing generates a focused summary of recent changes — active entities, recent facts, open commitments. Markdown or JSON.
Evaluation harness
Benchmark packs, shadow recall recording, CI delta gates. 672+ tests with CI enforcement.
Advanced (opt-in)
For users pushing the envelope. Each requires reading the relevant architecture doc.
Causal trajectories
Typed goal → action → observation → outcome chains.
Verified recall
Only surfaces memory boxes whose source memories still verify.
Semantic rule promotion
Promotes IF ... THEN rules from verified episodes.
Namespace policies
Principals, write rules, quota, and routing for multi-tenant installs.