RELATED 01
Engram
Gives coding agents a persistent, searchable project memory stored locally in SQLite
7.8kGitHub stars
5.7k/wknpm downloads
87popularity score
Remnic
Local, file-based long-term memory for coding agents and chat clients, served over HTTP MCP
By Joshua Warren (individual) TypeScript Local · Streamable HTTP Knowledge and memory
65 popularity score ★ 232 ↓ 6.6k/wk
Connect your client
Local (once it is running)
Add in one click
Opens the app, which asks before it adds the server.
Or add it by command or config
Claude CodeTerminal
claude mcp add --transport http remnic http://localhost:4318/mcpCursor~/.cursor/mcp.json
{
"mcpServers": {
"remnic": {
"url": "http://localhost:4318/mcp"
}
}
}VS Code.vscode/mcp.json
{
"servers": {
"remnic": {
"type": "http",
"url": "http://localhost:4318/mcp"
}
}
}Codex~/.codex/config.toml
[mcp_servers.remnic]
url = "http://localhost:4318/mcp"Gemini CLITerminal
gemini mcp add --scope user --transport http remnic http://localhost:4318/mcpSettings 6
REMNIC_AUTH_TOKEN |
Required: Bearer token clients must send; without it (or a token minted by `remnic token generate`) every request is refused. |
|---|---|
OPENAI_API_KEY |
Optional: Key for extracting memories with OpenAI; not needed with a local LLM, the OpenClaw gateway, or retrieval-only use. |
REMNIC_MEMORY_DIR |
Optional: Folder where the markdown memory files are kept. |
REMNIC_CONFIG_PATH |
Optional: Path to remnic.config.json. |
REMNIC_HOST |
Optional: Address to bind; 127.0.0.1 by default. |
REMNIC_PORT |
Optional: Port to listen on; 4318 by default. |
What it does
Remnic keeps an assistant's long-term memory on your own disk. Each memory is a markdown file with YAML frontmatter (facts, decisions, preferences, corrections, entities and a user profile) under a memory directory you choose, and a search index built from those files can be thrown away and rebuilt at any time. Recall ranks stored memories against the current context; a background step buffers conversation turns and uses a model of your choice to pull out durable facts and write them to disk.
The MCP side is a local HTTP service. You start it with `remnic-server` from @remnic/server or with `remnic daemon start` from @remnic/cli, and every MCP client (Claude Code, Codex CLI, Cursor, VS Code, Replit, Factory Droid and others) points at the same `/mcp` address with a bearer token, so one store is shared by all of them. Beyond store and recall it carries tools for correcting memories with an approval step, entity lookup, recall explanations, conversation-archive search, continuity and identity records, a small task and project board, and shared context between agents.
At v9.69.56 the source registers over 100 distinct tools, far more than the 44 listed in the connector guide, and several appear only when their feature is turned on in config (support passport, recall navigation, briefing, coding decisions, chat). The code advertises each tool as `remnic_<name>` (Anthropic tool names may not contain dots) and still accepts the `remnic.` and `engram.` forms the docs use (access-mcp-tool-names.ts).
What it needs
| Runtime | Node.js 22+ |
|---|
Node.js 22.12 or newer. A running Remnic server: `npx --package @remnic/server remnic-server` (or `npm install -g @remnic/cli`, then `remnic init` and `remnic daemon start`), with a bearer token set in REMNIC_AUTH_TOKEN. Writing new memories from conversations needs a model provider: an OpenAI API key, the OpenClaw gateway model chain, or a local LLM; retrieval-only use needs none. QMD on PATH is optional for the best hybrid search.
Limits
The standalone server has no stdio launch: clients connect over HTTP to the running service. With no auth token set it starts but rejects every request (packages/remnic-server/src/index.ts). It binds to 127.0.0.1 on port 4318 unless --host/--port or REMNIC_HOST/REMNIC_PORT are given. Namespaces, at-rest encryption and lossless transcript archiving are off by default. The Relay judge verifier is supported on Linux only and fails closed elsewhere. Tool names are moving from `engram.*` to `remnic_*` during the v1.x compatibility window.
Who it suits
Suits a developer who uses several coding agents or MCP clients and wants one memory they can read, grep, edit and keep in git, stored on their own machine rather than with a hosted memory service. It fits people willing to run a background service and supply a model for extraction (or a local LLM to keep everything on-device). Someone who wants a single stdio command with no daemon, or only needs a scratchpad for one client, will find the setup heavy. It is a community project, not from any of the agent vendors it connects to.
Tools 50 of 118
engram.recall- Find stored memories relevant to a query and return them as context
engram.recall_explain- Show the last recall snapshot for a session
engram.recall_xray- Run a recall and show which tier and filters produced each result
engram.recall_tier_explain- Explain the tier decision behind the last direct-answer recall
engram.observe- Send conversation turns into the archive and extraction pipeline
engram.memory_store- Save a memory directly
engram.memory_get- Fetch one memory by id
engram.memory_timeline- Show the history of one memory
engram.memory_search- Search memory files directly
engram.memory_correct_plan- Propose a correction to a stored memory
engram.memory_correct_apply- Apply an approved memory correction
engram.memory_feedback- Mark a recalled memory as relevant or not
engram.memory_promote- Move a memory to another lifecycle state
engram.memory_profile- Read the accumulated user profile
engram.memory_entities_list- List tracked people, projects and other entities
engram.entity_get- Fetch one entity by name
engram.memory_questions- List open questions gathered from conversations
engram.suggestion_submit- Queue a proposed memory for review
engram.review_queue_list- Show the current review queue
engram.review_list- List pending review items
engram.review_resolve- Accept or reject a review item
engram.set_coding_context- Tie a session to a project or branch so recall uses that scope
engram.lcm_search- Full-text search over archived conversation transcripts
engram.lcm_compaction_flush- Flush pending archive work before a client compacts its context
engram.lcm_compaction_record- Log a context compaction with token counts before and after
engram.extraction_force_flush- Drain a session's extraction buffer now
engram.day_summary- Build an end-of-day summary from today's memories
engram.briefing- Produce a briefing of active entities and open commitments, when enabled
engram.context_checkpoint- Save a session context checkpoint to disk
engram.action_confidence- Advise whether to ask, draft, act or refuse; read-only
engram.contradiction_scan_run- Scan the store for memories that contradict each other
engram.memory_governance_run- Run a governance pass over the store
engram.dreams_run- Run background consolidation of memories
engram.dreams_status- Show the state of background consolidation
engram.identity_anchor_get- Read the identity continuity anchor document
engram.identity_anchor_update- Merge new sections into the identity anchor
engram.continuity_incident_open- Open a continuity incident record
engram.continuity_incident_close- Close a continuity incident
engram.continuity_incident_list- List continuity incidents by state
engram.continuity_audit_generate- Write a weekly or monthly continuity audit report
engram.work_task- Create, update, move or delete tasks on the work board
engram.work_project- Manage work-board projects and link tasks to them
engram.work_board- Export or import a work-board snapshot
engram.peer_list- List known peers
engram.peer_set- Create or update a peer record
engram.peer_delete- Delete a peer record
engram.shared_context_write_output- Write an agent's output into shared context for other agents
engram.graph_snapshot- Return a snapshot of the entity graph
engram.memory_chat- Answer questions from the memory store, when chat is enabled
engram.profiling_report- Report recall and extraction timings, when profiling is on