Features

Everything Remnic does.

One memory store, every agent. This is the full capability list, grouped by three pillars: it works with every agent you run, it keeps your memory in plain files you own, and it lets you inspect, correct, and measure what it stores. Each card is marked shipped (on by default), opt-in (a config flag), or optional package (a separate install).

Works everywhere

Stop maintaining per-tool memory silos. One store on your disk, read and written by every agent you run and any MCP client, over stdio or HTTP.

Access surfaces

MCP stdio and HTTP Shipped

Over 100 MCP tools spanning recall, store, memory search, entity lookup, correction, temporal recall, X-ray provenance, daily briefing, work board, and LCM expansion. Every canonical remnic.* tool keeps an engram.* legacy alias (why the rename).

HTTP API Shipped

Bearer-token auth, binds to loopback by default. Recall, memories (CRUD), entities, observe, LCM, trust zones, review queue, and maintenance under /engram/v1/ (the compatibility prefix; see the rename note).

CLI Shipped

remnic ships 20+ commands: init, status, query, doctor, daemon, tree, onboard, curate, review, sync, dedup, connectors, space, benchmark, versions, taxonomy, enrich, binary, capsule, procedural.

Browser admin console Opt-in

Memory browser, recall debugger, trust-zone and review-queue panes, and a live memory-graph view at http://127.0.0.1:4318/remnic/ui/. Enable with server.adminConsoleEnabled.

ChatGPT access Shipped

Connect chatgpt.com in developer mode to a Remnic server you run, over MCP with OAuth 2.1. ChatGPT reads and writes the same store on your own machine as every other agent.

Work layer: tasks, projects, boards Shipped

Track tasks, projects, and boards through the remnic.work_task, remnic.work_project, and remnic.work_board tools. Work items stay out of memory extraction, so planning state never pollutes recall.

Multi-agent and identity

Namespaces Opt-in

Multi-agent memory isolation with principals, write rules, and session-based routing. Shared namespaces let agents that should collaborate see each other's memory.

Shared context Opt-in

Cross-agent intelligence for multi-agent setups. Agents share learnings through 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 can influence consolidation.

LLM routing

OpenAI Shipped

Set openaiApiKey and pick a model. GPT-5 family supported.

Local LLMs Shipped

Ollama, LM Studio, vLLM, or any OpenAI-compatible endpoint. The local-llm-heavy preset is tuned for offline operation.

Gateway model chain Opt-in

With modelSource: "gateway", route through OpenClaw agent personas with a primary plus fallbacks[]. Fireworks, then local, then OpenAI.

Per-call model selection Shipped

Different models for extraction, reranking, and consolidation. Cheap fast models on hot paths, smart models for synthesis.

Search backends

Six pluggable engines. All implement the same port, all configurable per deployment.

BackendTypeBest for
QMD (default)Hybrid BM25 + vector + rerankingBest recall quality
OramaEmbedded, pure JSZero native dependencies
LanceDBEmbedded, native ArrowLarge collections, fast vectors
MeilisearchServer-basedShared search across machines
RemoteHTTP RESTCustom search services
NoopNo-opExtraction-only deployments

Local-first and yours

Every memory is a plain markdown file with YAML frontmatter on your disk. No database is authoritative, the search index rebuilds from the files, and nothing phones home. MIT licensed, free forever, no telemetry.

At-rest encryption Opt-in

AES-256-GCM for memory files on disk, keyed by an Argon2id passphrase held only in memory. Protects a lost laptop or an exfiltrated backup. A running, unlocked daemon still serves plaintext to local clients, and the search index is not encrypted by this module.

Memory capsules (V2) Shipped

Portable, versioned archives in the V2 bundle format. Every file carries a SHA-256 checksum so imports are tamper-evident, and --encrypt seals the payload with AES-256-GCM. Fork, merge, and inspect capsules across machines and namespaces.

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.

Memory you can trust

Inspect why a memory surfaced, correct what is wrong, and measure whether recall actually helps. The intelligence that keeps the store sharp lives here, from the core loop to background consolidation.

User context model

User model Shipped

Preferences, goals, projects, constraints, priorities, communication style, risk tolerance, people, relationships, decisions, and definitions of good.

Scopes and boundaries Shipped

Personal, work, client, project, repo, tool, temporary, private, and do-not-use-outside-this-context boundaries.

Provenance metadata Shipped

Source, creation time, scope, retrieval reason, confidence, staleness, correction state, and current-context safety travel with every memory.

Action confidence Shipped

Advisory signals that help an agent decide whether to ask, draft, act, refuse, or escalate instead of spending attention carelessly.

Inspection surfaces Shipped

Recall, store, inspect, correct, forget, scope, and explain operations for MCP clients, plus a ChatGPT Apps-compatible memory inspector.

Core loop

Automatic extraction Shipped

Facts, decisions, preferences, and corrections are extracted from every conversation without manual effort.

Recall injection Shipped

Relevant memories are injected before each agent turn. Structural, not tool-based, so the agent cannot forget to recall.

Entity tracking Shipped

People, projects, tools, and companies tracked as structured entities with an addressable relationship graph.

Memory lifecycle Shipped

Active → validated → stale → archived. Old memories drop out of recall automatically, never from storage.

Episode and note model Shipped

Memories classified as time-specific episodes or stable beliefs. Episodes expire, notes persist.

Importance-gated writes Shipped

A local heuristic scores every memory (trivial, low, normal, high, critical) before it hits disk. Turn-level chatter never persists.

Inline source attribution Shipped

Compact provenance tags embed in fact bodies so citations survive prompt injection, copy and paste, and LLM quoting.

Cross-session recall Shipped

Memories from one session surface in another, across projects, machines, and agents.

Retrieval

Hybrid search Shipped

BM25 plus vector plus reranking via QMD. Query expansion and reranking happen inside QMD so the ranker sees more than the raw prompt.

Parallel specialized retrieval Shipped

Three retrievers run in parallel (DirectFact, Contextual, Temporal), weighted by confidence. Total latency is the max, not the sum.

Hard recall budget Shipped

recallBudgetChars caps total injected context. Per-section reservations stop any one source from hogging the budget.

Query-aware prefilter Shipped

Tag and temporal signals narrow the candidate set before hybrid search runs. Falls back to full search if the prefilter would over-trim.

Graph recall Opt-in

Entity-relationship graph traversal for causal and timeline queries alongside semantic recall.

Harmonic retrieval Opt-in

Blends abstraction nodes with cue-anchor matches, surfacing generalizations and specifics together.

Objective-state recall Opt-in

Surfaces file, process, and tool state snapshots alongside semantic memory, so the agent sees ambient state as well as stored knowledge.

Extraction and processing

Extraction priority order Shipped

Strict order: corrections, principles, preferences, commitments, decisions, relationships, entities, moments, skills, facts.

Local importance scoring Shipped

A zero-LLM regex engine scores every memory with explicit trivial-content short-circuits: greetings, one-word replies, emoji, anything under 10 characters.

Confidence tiers Shipped

Facts tagged explicit, implied, inferred, or speculative. Speculative memories auto-expire after 30 days unless confirmed.

Extraction judge Opt-in

LLM-as-judge post-extraction durability filter. Shadow mode calibrates before write-time gating turns on.

Semantic chunking Opt-in

Topic-boundary detection via sentence embeddings and cosine similarity with smoothing. An alternative to recursive chunking.

Enrichment pipeline Opt-in

Importance-tiered API spend for entity enrichment from external sources. Pluggable provider registry.

Consolidation and hygiene

Procedural memory Shipped

Recurring multi-step workflows are mined from causal trajectories and filed as procedure memories. On task-initiation prompts, recall injects a short relevant-procedures block. Enabled by default (procedural.enabled).

Dreams consolidation Shipped

Named background consolidation in three phases: light sleep scores recent activity, REM synthesizes and de-duplicates across sessions, deep sleep promotes durable findings and migrates cold material. Light sleep runs by default; REM and deep sleep are opt-in.

Exact-hash dedup Shipped

A content-hash index at every write. Facts whose content already exists never write again. Chunked memories register their parent content in the same index.

Fuzzy duplicate scanner Shipped

remnic dedup runs Jaccard plus substring-containment across categories and flags pairs at configurable thresholds for merge or review.

Contradiction detection Shipped

Negation-aware pairwise scan. Finds "X is true" against "X is not true" and surfaces high and medium severity conflicts.

Memory cache Shipped

A process-level singleton cache for readAllMemories(). Turns 15s disk scans into cache hits shared across all sessions.

Semantic consolidation Opt-in

Clusters of similar memories detected via cosine similarity. An LLM synthesizes a canonical version. Originals archived with full provenance, not deleted.

LLM consolidation Opt-in

A scheduled pass asks the model to ADD, MERGE, UPDATE, INVALIDATE, or SKIP each new memory against existing ones.

Pattern reinforcement Opt-in

Recurring preferences, facts, and decisions cluster into one canonical primitive with a reinforcement count, so repeated signals outrank one-off mentions. A separate opt-in weights reinforced primitives higher in recall.

Provenance and trust zones

Recall X-ray Shipped

Per-result provenance for every recall: which retrieval tier produced each memory, its score components, and why it surfaced. Inspect it over the API, the MCP tool, or the browser console.

Trust zones Opt-in

A three-tier model: quarantine → working → trusted. New memories land in quarantine and promote based on provenance and corroboration, with a guarded path that blocks direct quarantine-to-trusted jumps.

Corroboration scoring Opt-in

Memories gain trust when independent sources confirm the same fact. Corroboration is automatic and auditable.

Poisoning defense Opt-in

Deterministic provenance trust scoring. Risky promotions into trusted require anchored provenance and at least one independent corroborating record.

OAI-mem-citation blocks Opt-in

Recall emits <oai-mem-citation> blocks matching the Codex citation format for memory attribution and usage tracking.

Trust-zone demo dataset Shipped

A buyer-friendly enterprise scenario via remnic trust-zone-demo-seed. Never seeded automatically.

Lossless context management (LCM)

When an agent hits its context window, the runtime silently compresses old messages and that context is gone. LCM archives every message into a local SQLite database and builds a hierarchical summary tree alongside it.

Proactive archiving Opt-in

Every message is indexed with full-text search before native compaction can discard it.

Hierarchical summaries Opt-in

Leaf summaries cover about 8 turns, depth-1 about 32, depth-2 about 128. The tree is built lazily on demand.

Fresh tail protection Opt-in

Recent turns always use the most detailed leaf-level summaries. Older turns get compressed further.

Three-level summarization Opt-in

Normal LLM summary, aggressive bullet compression, deterministic truncation. Guaranteed convergence, no LLM needed for the fallback.

MCP expansion tools Opt-in

Agents can search, describe, or expand any part of conversation history on demand.

Configurable retention Opt-in

Raw messages retained for the configured window (default 90 days) and queryable via full-text search.

Organization

MECE taxonomy Opt-in

A mutually exclusive, collectively exhaustive knowledge directory with a resolver decision tree for deterministic categorization.

Memory boxes Opt-in

Groups related memories into topic-windowed episodes that stay internally consistent.

Memory OS (progressive)

Start with zero config and enable features as needs grow. Presets (conservative, balanced, research-max, local-llm-heavy) jump you to a recommended level.

Compounding Opt-in

Weekly synthesis surfaces patterns and recurring mistakes.

Native knowledge Opt-in

Search curated markdown (workspace docs, Obsidian vaults) without extracting into memory.

Behavior loop tuning Opt-in

Runtime self-tuning of extraction and recall parameters based on outcomes.

Creation memory Opt-in

A work-product ledger tracking agent outputs and their durability.

Commitment lifecycle Opt-in

Tracks promises, deadlines, and obligations with review surfaces.

Resume bundles Opt-in

Crash-recovery context for interrupted sessions.

Utility learning Opt-in

Learns promotion and ranking weights from downstream outcomes.

Coding intelligence

Coding graph Optional package

The @remnic/coding-graph package indexes a repository into a SQLite knowledge graph: web-tree-sitter parsing, structural queries, blast-radius analysis, and dead-code detection. Gated by codingKnowledge.codegraphTools and installed separately.

Advanced

Causal trajectories Opt-in

Typed goal → action → observation → outcome chains.

Verified recall Opt-in

Only surfaces memory boxes whose source memories still verify.

Semantic rule promotion Opt-in

Promotes IF-THEN rules from verified episodes.

Namespace policies Opt-in

Principals, write rules, quota, and routing for multi-tenant installs.

Inspect and measure

Daily briefing Shipped

remnic briefing generates a focused summary of recent changes: active entities, recent facts, open commitments. Markdown or JSON.

Memory evals Shipped

Tests for repeated-context reduction, unnecessary clarification, retrieval quality, stale-memory harm, scope respect, and output improvement.

Evaluation harness Shipped

Benchmark packs, shadow recall recording, and CI delta gates with enforcement.