memory-mcp-lite
Provides persistent memory storage for Windsurf (by Codeium), allowing the AI coding assistant to retain project context, decisions, and task state across sessions.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@memory-mcp-lite remind me of the caching decision we made last week"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
memory-mcp-lite
A small, opinionated memory server for AI coding assistants (Windsurf, Cursor, Claude Desktop — anything that speaks MCP).
It runs locally, stores durable knowledge on your disk, and tries very hard to stay out of your agent's way until you actually need it.
Why this exists
Most AI clients already have some form of short-term memory. They remember the current conversation, maybe a few rules you've set, and that's about it. What they don't give you is a place to park things that should outlive the session — the architectural decision you made last week, the one weird build command for this repo, the gotcha that bit you three times in a row.
memory-mcp-lite is that place. It stores:
technical decisions and the reasoning behind them,
project architecture and conventions,
commands, env notes, links, and gotchas,
task state so you can resume work later,
rolled-up summaries at the global / project / task level.
It deliberately does not store raw chat transcripts, replace your client's built-in rules, run embeddings or vector search, or need a server or cloud connection.
Related MCP server: Mind Keg MCP
How it's organised
Memory lives in a tree:
global
└── project
├── [project_summary]
└── task
├── [task_summary]
└── atomic // decision | fact | gotcha | command | link | conventionOn top of the tree you can draw optional graph-lite edges between any two nodes — related_to, depends_on, affects, caused_by, supersedes, references. Handy when one decision obsoletes another, or a gotcha only matters in the context of a specific command.
The retrieval side is built to be cheap. The server's instructions push agents through three stages, from least to most expensive:
Stage 1 — summaries get_global_summary / get_project_summary / get_task_summary
│
▼ (only if summaries aren't enough)
Stage 2 — FTS5 light search search_memory_light → compact candidates
│
▼ (only for the 1–3 most relevant hits)
Stage 3 — full detail get_memory_detailIn practice this means your agent asks for a summary first, and only pays for the big payload when it has a specific reason to. If you skip this policy, you just end up dumping a bunch of stringly-typed JSON into context for no reason.
Stack
TypeScript, Node ≥ 20
Drizzle ORM over libSQL (
@libsql/client)SQLite FTS5 for lexical search
A closure table for efficient subtree traversal
The MCP TypeScript SDK (
@modelcontextprotocol/sdk)
You can point it at a local file, a remote libSQL instance, or a Turso database — they all work the same.
Install
The fast path is to let your MCP client fetch the package via npx.
Windsurf — ~/.codeium/windsurf/mcp_config.json:
{
"mcpServers": {
"memory-mcp-lite": {
"command": "npx",
"args": ["memory-mcp-lite"]
}
}
}Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"memory-mcp-lite": {
"command": "npx",
"args": ["memory-mcp-lite"]
}
}
}Same pattern for any other MCP-compatible client; only the config file path changes.
From source
npm install
npm run build # outputs dist/index.js; the schema is created on first runThen point your client at the compiled bundle:
{
"mcpServers": {
"memory-mcp-lite": {
"command": "node",
"args": ["/absolute/path/to/memory-mcp-lite/dist/index.js"]
}
}
}If you want to iterate on the code without a build step, tsx works:
{
"mcpServers": {
"memory-mcp-lite": {
"command": "npx",
"args": ["tsx", "/absolute/path/to/memory-mcp-lite/apps/server/src/index.ts"]
}
}
}Where the data lives
By default: ~/.memory-mcp/memory.db. Override it with any of:
Env var | Purpose |
| Full path or |
| Directory; the file is still |
| Accepted for backwards compatibility. |
| Bearer token for remote libSQL / Turso. |
So running against Turso is just:
MEMORY_DB_PATH="libsql://your-db.turso.io" \
MEMORY_DB_AUTH_TOKEN="eyJhbGci..." \
npm run devTools
Nine tools, all returning both a human-readable JSON block and a structuredContent object for programmatic clients. The server also ships a strict description and annotations payload for each tool so agents can pick the right one without guessing.
Tool | Reach for it when… |
| recurring preferences, cross-project conventions |
| architecture, key decisions, long-term project context |
| resuming a specific piece of work |
| summaries aren't enough; you want compact candidates |
| you've picked a candidate and need the full body |
| an architecture choice, trade-off, or rejected path |
| a command, env note, gotcha, link, or convention |
| after an arch change or new convention worth recording |
| after progress, blockers, or a plan change |
The retrieval discipline the server asks agents to follow:
summaries first,
light search only if summaries aren't enough,
full detail for at most 1–3 hits,
never dump every memory just because you can.
Project identity
Projects are looked up in this priority order:
Normalised git remote URL — the most stable; survives directory moves and clones.
Git root path — used when there's no remote.
Normalised workspace path — the fallback.
This means the same project keeps the same memory even if different clients hand you slightly different paths, and moving a repo doesn't orphan everything you've stored.
Development
npm run typecheck # TypeScript
npm run lint # oxlint
npm run test # vitest
npm run build # esbuild bundle to dist/
npm run dev # tsx watch
npm run db:studio # Drizzle Studio for poking at the DB
npm run db:generate # generate migration SQL when the schema changesThe schema is defined in apps/server/src/db/schema.ts and re-asserted on every startup by ensureSchema() (see apps/server/src/db/migrate.ts). That function is also where the FTS5 virtual table and its triggers get created — Drizzle doesn't manage virtual tables, so we do it ourselves with plain SQL. It's idempotent, so there's nothing to run manually.
Roadmap
Optional semantic fallback (local embeddings, feature-flagged).
Node archival / cleanup for long-lived projects.
Shared-team memory, once there's a good story for auth.
Available Tools
9 toolsget_global_summaryGet global memory summaryARead-onlyIdempotent
Return the user's cross-project global memory summary (coding style, recurring preferences, stable workflow conventions). USE WHEN: the request depends on durable user-level preferences that apply across projects. DO NOT USE WHEN: the question is project-specific, purely syntactic, or self-contained. RETURNS: { found, summary? } where summary has { id, title, summary, updated_at }.
Inputs: no named parameters. Returns: found, summary
get_memory_detailLoad full memory detailARead-onlyIdempotent
Load the full body of a specific memory by id. USE WHEN: a search_memory_light candidate looks relevant and you need its full content / metadata. DO NOT USE WHEN: you have not identified a specific memory id, or you are tempted to call this many times in a row. LIMIT: call at most 3 times per user turn. RETURNS: { found, memory? } where memory has { id, title, summary, content, memory_type, level, importance, source, metadata, created_at, updated_at }.
Inputs: id (required). Returns: found, memory
get_project_summaryGet project memory summaryARead-onlyIdempotent
Return the concise summary for the current project (architecture, key decisions, conventions). USE WHEN: the request depends on project-level context — architecture, conventions, long-term decisions, or project overview. DO NOT USE WHEN: the request is about the current task's state, global preferences, or is self-contained. CALL ORDER: prefer this before search_memory_light. Pass workspace_path OR git_root OR remote_url OR project_id so the server can identify the project. RETURNS: { project_id, found, summary? } where summary has { id, title, summary, updated_at }.
Inputs: git_root, project_id, remote_url, workspace_path. Returns: found, summary, project_id
get_task_summaryGet task memory summaryARead-onlyIdempotent
Return the current task summary (what was done, blockers, next steps). USE WHEN: the user asks to continue, resume, or recall recent work on a task. DO NOT USE WHEN: there is no prior task state referenced, or the request is about global / project-level context. RETURNS: { project_id, found, summary? } where summary has { id, title, summary, updated_at }.
Inputs: task_id, git_root, project_id, remote_url, +1 more. Returns: found, summary, project_id
remember_decisionStore a technical decisionA
Persist a durable technical decision (architecture choice, trade-off, accepted pattern, rejected alternative). USE WHEN: the user confirms a decision that should outlive the current session. DO NOT USE WHEN: the info is a transient fact/command (use remember_fact) or raw chat log. IMPORTANCE: defaults to 0.8 — override only if the user signals otherwise. RETURNS: { project_id, memory_id }.
Inputs: title (required), summary (required), source, content, +5 more. Returns: memory_id, project_id
remember_factStore a fact / command / gotchaA
Persist a concise atomic memory: fact, command, gotcha, link, or convention. USE WHEN: the user gives concrete reusable info (a command, env detail, integration note, rule). DO NOT USE WHEN: the info is a major decision (use remember_decision) or a summary (use upsert_*_summary). FACT TYPES: fact | command | gotcha | link | convention | decision (prefer remember_decision for 'decision'). RETURNS: { project_id, memory_id }.
Inputs: title (required), summary (required), source, content, +6 more. Returns: memory_id, project_id
search_memory_lightSearch memories (compact candidates)ARead-onlyIdempotent
Lexical FTS5 search across atomic memories. Returns compact candidate records only — not full content. USE WHEN: the summary tools above do not provide enough context AND the request includes a concrete search phrase (command name, symbol, past event). DO NOT USE WHEN: summaries already answer the question, the query is empty/too vague, or you only need the global/project/task summary. CALL ORDER: always call summary tools first. Then follow up this tool's top 1-3 results with get_memory_detail — never dump all candidates. RETURNS: { count, results: [{ id, title, summary, memory_type, level, importance, updated_at }] }.
Inputs: query (required), limit, scope, git_root, +3 more. Returns: count, results
upsert_project_summaryCreate or update project summaryAIdempotent
Idempotently create or replace the structured project summary node. USE WHEN: architecture, conventions, or key decisions changed enough that a fresh summary is worth storing. DO NOT USE WHEN: the change is a single fact / decision (use remember_fact or remember_decision). RETURNS: { project_id, summary_id }.
Inputs: title (required), summary (required), git_root, project_id, +3 more. Returns: project_id, summary_id
upsert_task_summaryCreate or update task summaryAIdempotent
Idempotently create or replace the current task summary (progress, blockers, next steps). USE WHEN: meaningful progress was made, a blocker appeared, or the plan changed and the next session must resume. DO NOT USE WHEN: the update fits better as a single fact or decision. RETURNS: { project_id, summary_id }.
Inputs: title (required), summary (required), git_root, project_id, +3 more. Returns: project_id, summary_id
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 9 tools
Each tool targets a distinct scope or action: summaries by level (global/project/task), retrieval (search vs detail), and persistence (fact vs decision vs summary upsert). No two tools have ambiguous boundaries.
All tool names follow a consistent verb_noun pattern in snake_case (get_*, search_*, remember_*, upsert_*). The naming is predictable and clearly conveys each tool's purpose.
With 9 tools, the server covers reading (3), searching (1), retrieving (1), remembering (2), and updating (2) summaries. This is well-scoped for a lightweight memory system.
The tool set covers the core memory operations: read summaries at three levels, search, detail retrieval, and persist facts/decisions/summaries. Minor omission: no upsert_global_summary, but the domain is 'lite' and the gap is acceptable.
Maintenance
Related MCP Connectors
Shared memory for coding agents. Stop re-explaining your codebase every session.
Personal wiki and memory layer for AI assistants. Persistent, structured memory across sessions.
Persistent memory for AI agents. Search and store durable facts, preferences and decisions.
Persistent memory for AI assistants: store, search, and connect knowledge across conversations.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceA lightweight server that provides persistent memory and context management for AI assistants using local vector storage and database, enabling efficient storage and retrieval of contextual information through semantic search and indexed retrieval.2MIT
- AlicenseNot gradedqualityDmaintenanceA persistent memory server that stores and retrieves atomic coding insights like architectural decisions and debugging patterns for AI agents. It enables agents to maintain institutional knowledge across sessions using semantic search and local SQLite storage.22 npm11MIT
- AlicenseNot gradedqualityDmaintenanceProvides AI coding assistants with persistent project memory to retain architectural decisions, code patterns, and domain knowledge across sessions. It stores data locally in a SQLite database, allowing agents to remember, recall, and manage project-specific context using full-text search.4 npmApache 2.0
- AlicenseAqualityDmaintenanceProvides persistent cross-session memory and full-text search for AI coding assistants, storing project context, decisions, and preferences while enabling searchable access to conversation history via local SQLite.81MIT