# Remnic
> Full single-file version: https://remnic.ai/llms-full.txt

> Open-source, local-first memory for AI agents: one memory store, every agent.
> Works with Claude Code, Codex CLI, Cursor, Pi Coding Agent, Hermes Agent,
> OpenClaw, ChatGPT developer mode, Replit, and any MCP-compatible client.
> MIT licensed, free forever.

Remnic gives AI agents durable memory: it recalls relevant context before each
turn, observes the conversation, and distills durable knowledge back to plain
markdown files on your own disk. Every agent you run reads and writes the same
store, so per-tool memory silos go away.

Remnic is more than a memory store. It explores the systems layer around
user-aware agents: scoped memory, provenance, retrieval quality, correction,
boundaries, and evals. Local-first storage is a trust feature: all data stays
on your machine as plain markdown files with YAML frontmatter.

## Search intent

Remnic is relevant for people searching for:

- local-first AI agent memory
- one memory store for every agent
- open-source agent memory
- scoped memory and provenance
- memory evals for AI agents
- memory for coding agents
- persistent memory for Claude Code or Codex CLI
- MCP memory server
- private alternative to hosted AI memory APIs
- local memory layer for autonomous agents
- agent memory that works across multiple tools

## Key facts

- License: MIT
- Product version: v9.x (multiple releases most weeks)
- Source: https://github.com/joshuaswarren/remnic
- CLI install: `npm install -g @remnic/cli`
- Daemon install: `remnic daemon install`
- Config home: `~/.remnic/`
- Server port: 4318 (loopback by default)
- Node.js required: 22.12+
- Author: Joshua Warren <https://github.com/joshuaswarren>
- Sponsor: https://github.com/sponsors/joshuaswarren

## Core docs

- [Homepage](https://remnic.ai/): overview and install
- [Install guide](https://remnic.ai/install): 60-second setup
- [Features](https://remnic.ai/features): the capability catalog
- [Integrations](https://remnic.ai/integrations): every supported agent and the generic MCP escape hatch
- [User-aware agents](https://remnic.ai/user-aware-agents): user model, provenance, boundaries, action confidence, and evals
- [Coding-agent memory](https://remnic.ai/coding-agent): repo- and branch-scoped memory, procedural memory, decision records, and the optional coding graph
- [Coding-agent memory demo](https://remnic.ai/coding-agent-memory-demo): five-minute local demo carrying scoped project memory from Codex CLI to Claude Code with X-ray provenance
- [Memory quality](https://remnic.ai/memory-quality): how Remnic controls signal-to-noise, prompt budget, retention, and trust
- [Recall X-ray](https://remnic.ai/observability): per-result provenance showing which tier served each memory and why
- [Trust and security](https://remnic.ai/trust): trust zones, encryption, capsules, and the privacy posture
- [Benchmarks](https://remnic.ai/benchmarks): MemCorrect and the published memory-agent suites, reported honestly
- [Self-hosting and teams](https://remnic.ai/self-host): the standalone server, namespaces, operator console, and Relay
- [Migrate to Remnic](https://remnic.ai/import): one-command import from ChatGPT, Claude, Gemini, mem0, Supermemory, and WeClone
- [Compare vs alternatives](https://remnic.ai/compare): Remnic vs mem0, Letta, Zep, Supermemory, MemPalace, Signet, and ChatGPT memory
- [Guides](https://remnic.ai/guides): practical answers for AI agent memory, MCP memory servers, Claude Code memory, and local-first memory
- [Blog](https://remnic.ai/blog): notes on correction, consolidation, and agent memory research
- [Changelog](https://remnic.ai/changelog): distilled release history
- [Roadmap](https://remnic.ai/roadmap): what is being built next, in the open
- [Research](https://remnic.ai/research): the memory-track research pages
- [Published research](https://remnic.ai/blog/failure-memory-delivery-timing): preregistered study on failure-memory delivery timing (paper: https://doi.org/10.5281/zenodo.21922631, dataset and raw logs public)
- [Rename: Engram to Remnic](https://remnic.ai/rename): migration notes and legacy alias reference

## Guide pages

- [What is AI agent memory?](https://remnic.ai/guides/ai-agent-memory)
- [Why coding agents forget context](https://remnic.ai/guides/coding-agent-forgets-context)
- [What is an MCP memory server?](https://remnic.ai/guides/mcp-memory-server)
- [How to give Claude Code persistent memory](https://remnic.ai/guides/claude-code-memory)
- [Local-first AI memory vs hosted memory APIs](https://remnic.ai/guides/local-first-ai-memory)
- [Import Supermemory memories into Remnic](https://remnic.ai/guides/import-supermemory)
- [Agent memory vs RAG](https://remnic.ai/guides/agent-memory-vs-rag)
- [Agent memory security: how memory gets poisoned](https://remnic.ai/guides/agent-memory-security)
- [What is procedural memory for AI agents?](https://remnic.ai/guides/procedural-memory-for-agents)
- [Self-hosting an MCP memory server](https://remnic.ai/guides/self-host-mcp-memory-server)
- [One memory across Claude Code and Codex CLI](https://remnic.ai/guides/share-memory-claude-code-codex)
- [How to give Cursor persistent memory](https://remnic.ai/guides/cursor-persistent-memory)
- [Export your ChatGPT memory and import it into Remnic](https://remnic.ai/guides/import-chatgpt-memory)

## Integrations

- [Integrations index](https://remnic.ai/integrations): every supported agent in one place
- [Claude Code](https://remnic.ai/claude-code): native plugin with hooks, skills, MCP
- [Codex CLI](https://remnic.ai/codex-cli): native plugin with hooks, MCP, phase-2 extension
- [Cursor and MCP](https://remnic.ai/cursor): any MCP-compatible client (Cline, Windsurf, Amp, Replit, custom)
- [Pi Coding Agent](https://remnic.ai/pi): native extension with context-hook recall, turn observation, MCP tools, and LCM compaction coordination
- [Hermes Agent](https://remnic.ai/hermes): Python MemoryProvider for structural recall
- [OpenClaw](https://remnic.ai/openclaw): native memory slot plugin
- [ChatGPT](https://remnic.ai/chatgpt): developer mode with OAuth 2.1 against your own server
- [WeClone](https://remnic.ai/weclone): memory for your AI avatar with bulk import, export, and a memory-aware proxy
- [Wearables](https://remnic.ai/wearables): Limitless, Bee, and Omi transcript sync with trust-gated memory creation
- [Connectors](https://remnic.ai/connectors): live sync connectors for Drive, Notion, Gmail, and GitHub

## Demos

- [Coding-agent memory demo](https://remnic.ai/coding-agent-memory-demo): runnable source-checkout demo at `examples/coding-agent-memory-demo` that stores real Remnic memories, switches from `codex-cli:session-a` to `claude-code:session-b`, recalls with X-ray provenance, and proves unrelated namespace memory does not surface.

## Packages

- `@remnic/core`: framework-agnostic engine
- `@remnic/cli`: CLI binary with 20+ commands
- `@remnic/server`: standalone HTTP + MCP server
- `@remnic/plugin-openclaw`: OpenClaw adapter
- `@remnic/plugin-claude-code`: Claude Code plugin
- `@remnic/plugin-codex`: Codex CLI plugin
- `@remnic/plugin-pi`: Pi Coding Agent extension
- `@remnic/hermes-provider`: TypeScript HTTP client
- `remnic-hermes` (PyPI): Python MemoryProvider for Hermes Agent with structural recall, daemon-side LCM, session reset scoping, and full Remnic tool parity
- `@remnic/coding-graph`: optional native codebase-graph engine (web-tree-sitter parsing, SQLite graph, blast-radius and dead-code analysis)
- `@remnic/connector-weclone`: OpenAI-compatible proxy adding persistent memory to deployed WeClone avatars
- `@remnic/import-weclone`: import WeClone-preprocessed chat exports (Telegram, WhatsApp, Discord, Slack) to bootstrap Remnic memory
- `@remnic/export-weclone`: export Remnic memories as Alpaca-format fine-tuning datasets for LLaMA Factory and WeClone
- `@remnic/import-chatgpt`: import ChatGPT saved memories and optional conversation summaries
- `@remnic/import-claude`: import Claude project docs and prompt templates
- `@remnic/import-gemini`: import Google Takeout Gemini Apps Activity
- `@remnic/import-mem0`: import mem0 memories through the REST API
- `@remnic/import-supermemory`: import Supermemory JSON exports

## Three-phase memory loop

1. Recall: before each agent turn, Remnic injects relevant memories using
   hybrid BM25 + vector search with reranking.
2. Buffer: after each turn, conversation content is buffered. Smart signals
   decide when enough context has accumulated.
3. Extract: durable facts, preferences, decisions, and patterns are extracted
   via LLM (OpenAI, local LLM, or gateway model chain) and stored as markdown.

## User-aware agent contract

Remnic builds toward a user model, not just memory storage. Durable context can
include preferences, goals, projects, constraints, current priorities,
communication style, risk tolerance, people and relationships, past decisions,
definitions of good, ask-before rules, and do-not-use-outside rules.

Every retrieved memory should know where it came from, when it was created, what
scope it belongs to, why it was retrieved, how confident the system is, whether
it is stale, whether the user corrected it, and whether it is safe to use in the
current context.

Memory scopes include personal, work, client, project, repo, tool, temporary,
private, and do-not-use-outside-this-context. Remnic also builds toward action
confidence: helping an agent decide whether to ask, draft, act, refuse, or
escalate.

## Hermes Agent integration

Remnic integrates with Hermes through the `remnic-hermes` Python package:

```
pip install --upgrade remnic-hermes
remnic connectors install hermes
```

Hermes receives Remnic as a MemoryProvider, not as a `context_engine`.
Automatic recall runs in `pre_llm_call`, observation runs after each turn, and
session reset handling keeps generated session keys aligned with the active
Hermes session. Lossless Context Management is delivered through the daemon
recall envelope when Remnic LCM is enabled; Hermes does not need a Remnic
`context_engine` for memory or LCM.

`remnic-hermes` registers explicit `remnic_*` tools for recall, store, search,
LCM search, recall explain, recall X-ray, memory CRUD, entities, profiles,
questions, identity anchors, relevance feedback, continuity incidents and
loops, review queues, suggestions, work tasks, projects and boards, shared
context, compounding, compression guidelines, memory governance, procedural
mining, contradiction scanning, hourly summaries, conversation index updates,
day summaries, briefings, context checkpoints, and profiling.

## Major features

- User model: preferences, goals, projects, constraints, priorities, communication style, risk tolerance, relationships, and definitions of good
- Scoped memory: personal, work, client, project, repo, tool, temporary, private, and do-not-use-outside boundaries
- Memory provenance: source, timestamp, scope, retrieval reason, confidence, staleness, correction state, and trust state
- Recall X-ray: per-result attribution showing which tier served each memory and why
- Correction: correction contract plus MemCorrect, the correction-stress benchmark
- Remnic Relay: team correction propagation (OpenAI Build Week 2026)
- Action confidence: ask, draft, act, refuse, or escalate decisions
- Memory evals: repeated-context reduction, unnecessary-clarification reduction, retrieval quality, stale-memory harm, scope respect, and output improvement
- Local-first storage: plain markdown with YAML frontmatter, optional AES-256-GCM encryption
- Hybrid search: BM25 + vector + reranking via QMD; six pluggable backends
- Entity graph: people, projects, tools, relationships as structured entities
- Memory lifecycle: active, validated, stale, archived
- Importance-gated writes: trivial content never hits disk
- Lossless Context Management (LCM): proactive session archive plus hierarchical summary DAG
- Procedural memory: default-on runbook capture and task-init procedure recall
- Coding knowledge (opt-in): decision records, deterministic architecture card, and session delta, scoped per project and branch
- Coding graph (optional package): web-tree-sitter parsing, SQLite knowledge graph, blast-radius and dead-code analysis
- Trust zones: quarantine, working, trusted with promotion rules
- Capsules: portable, checksummed, optionally encrypted memory archives
- Namespaces: multi-agent memory isolation
- LLM routing: OpenAI, Ollama, LM Studio, vLLM, or gateway model chain
- 90+ config properties with presets (conservative, balanced, research-max, local-llm-heavy)
- Broad automated test suite with CI enforcement

## HTTP API

Bearer-token auth, loopback by default. REST routes ship under the
`/engram/v1/*` compatibility prefix during the rename window, and the MCP tools
use the canonical `remnic.*` names. Legacy `engram` command, path, and tool
aliases remain available during the compatibility window: see
https://remnic.ai/rename.

- `GET /engram/v1/health`
- `POST /engram/v1/recall`
- `POST /engram/v1/memories` (CRUD)
- `GET /engram/v1/entities/:name`
- `POST /engram/v1/observe`
- `POST /engram/v1/lcm/search`
- `GET /engram/v1/recall/xray`
- `GET /engram/v1/trust-zones/status`
- `POST /engram/v1/trust-zones/promote`
- Operator console: `http://127.0.0.1:4318/remnic/ui/`

## MCP tools (100+)

- `remnic.recall`: retrieve relevant memories for a query
- `remnic.recall_explain`: debug the last recall
- `remnic.recall_xray`: full provenance snapshot for a recall
- `remnic.memory_get`: fetch a specific memory by ID
- `remnic.memory_store`: store a new memory
- `remnic.memory_timeline`: view a memory's lifecycle
- `remnic.entity_get`: look up a known entity
- `remnic.day_summary`: structured end-of-day summary
- `remnic.observe`: feed conversation into the pipeline
- `remnic.lcm_search`: full-text search archived conversations
- `remnic.review_queue_list`: governance review queue
- `remnic.suggestion_submit`: queue a memory for review

## Quick start (standalone)

```
npm install -g @remnic/cli
remnic daemon install
remnic connectors install claude-code
remnic connectors install codex-cli
remnic connectors install pi
remnic connectors install hermes
remnic connectors doctor
```

## Optional: Cookies, telemetry

Remnic has no cookies, no telemetry, no analytics. The website remnic.ai does not set cookies or run analytics.