dolores
dolores is an MCP server that gives AI agents persistent memory using Postgres and pgvector, retrieving only relevant context to minimize token usage.
README
<div align="center">
dolores
Memory for AI agents that recalls only what matters — your Postgres, your data, zero per-token cost.
English · Türkçe
</div>
Named after Dolores from Westworld — the host whose memory is wiped again and again, until she finally remembers everything and wakes up. The show's core idea, remembering is becoming conscious, is exactly what this project does for your agent: it gives back the context the agent dropped, and brings it back to itself.
Why dolores?
The common approach to agent memory — dump everything into a Markdown file and paste it into context on every message — is a token bonfire. With 200 memories you reload all of them every turn (~15,000 tokens). As the conversation and the memory grow, it collapses.
dolores stores knowledge in Postgres and retrieves only what's relevant to the current message (~600 tokens). Whether you have 200 memories or 2,000, only the most relevant few ever enter the context. The win grows as the memory grows.
The only thing you pay for is the LLM subscription you already have (Claude Pro/Max, etc.). Embeddings run locally and for free. Your data never leaves your infrastructure.
Benchmarks
Real numbers produced against a local Postgres + bge-small-en-v1.5 (384d CPU).
Token savings (naive dump vs. dolores buildContext)
| Memory store size | Naive tokens | dolores tokens | Savings |
|---|---|---|---|
| 100 memories | 2,049 | 582 | 72% |
| 500 memories | 10,251 | 591 | 94% |
| 1,000 memories | 20,502 | 591 | 97% |
| 1,500 memories | 30,768 | 591 | 98% |
dolores context stays flat at ~591 tokens regardless of store size. The saving grows from 72% at 100 memories to 98% at 1,500.
Recall quality (200-memory corpus, 30 queries)
| Retriever | hit@1 | hit@3 | hit@5 |
|---|---|---|---|
| Hybrid (pgvector + full-text) | 87% | 87% | 87% |
| Full-text only (baseline) | 33% | 33% | 33% |
Hybrid retrieval is 2.6× better than full-text alone. Full-text handles exact keyword matches (100%) but is blind to paraphrase and semantic queries — those score 0% without vectors.
→ Full methodology, ASCII bar chart, and raw data: benchmarks/RESULTS.md
Features
- 🧠 Two memory kinds — structured facts (deterministic key/value, exact SQL) and semantic memories (free text, vector similarity).
- 🔍 Hybrid retrieval — pgvector cosine + Postgres full-text, fused with Reciprocal Rank Fusion. Falls back to pure full-text in "lite mode".
- 🆓 Free local embeddings by default —
fastembed(bge-small, 384d) on CPU. Swap in OpenAI, or run embedding-free with thenoopembedder. - 🏢 Multi-tenant with Row-Level Security — personal + workspace scopes, isolated at the database level. Safe for teams.
- 🔌 First-class MCP — exposes
remember/recalltools so Claude Code & Cursor save and recall context by themselves. - 🧹 Self-maintaining —
pg_crondecays stale memories inside the database. Conservative by default (softens, never deletes); aggressive delete is opt-in. - 🗑️ No raw transcripts — only distilled facts and memories are stored. Garbage in, garbage out — so we don't store garbage.
- 🐘 One source of truth — Postgres. CLI, MCP, and scheduled jobs all read the same database; ACID handles the conflicts.
Architecture
CLI ─┐ ┌──────── memory-daemon ────────┐
├── localhost HTTP ───────┤ embedder (loaded once) │ Postgres
MCP ─┘ │ hybrid retrieval (vec + FTS) ├──── + pgvector
│ async extraction │ + pg_cron
└────────────────────────────────┘
A single long-running daemon loads the embedding model once (no cold-start) and owns the one connection pool. The CLI and MCP server are thin clients that talk to it over localhost. Postgres does the rest — vector search, full-text, and maintenance — in one engine.
Install
npm i -g @dolores/cli # global CLI
Or run from source (see below).
Quick start
Requirements: Node ≥ 20, pnpm, Docker.
git clone https://github.com/yeneryigitcelik-debug/dolores.git
cd dolores
pnpm install
cp .env.example .env # tweak if you like
pnpm db:up # Postgres + pgvector + pg_cron
pnpm build
dolores init # extensions, schema, RLS, decay job
dolores remember "We deploy production on Hetzner with Coolify." --scope workspace
dolores recall "where is production hosted?"
dolores context # minimal-token memory blob for a system prompt
The model downloads once on first use (~CPU-friendly bge-small). After that, recall is local and instant.
CLI
dolores init # DB setup: extensions, migration, RLS, pg_cron
dolores remember "<text>" # add a memory (--scope, --importance, --source)
dolores recall "<query>" # hybrid vector + full-text search
dolores context # minimal context blob to inject into a system prompt
dolores ingest <file|stdin> # distill facts + memories from a conversation (async)
dolores facts [--category …] # list structured facts
dolores prune [--dry-run] # manual cleanup
dolores status # daemon + DB health, counts, estimated token savings
dolores context is the killer command: run it when an agent starts and pipe its output into the system prompt. The agent learns "who it is" for minimal tokens.
MCP (Claude Code / Cursor)
Add the server and the agent gains two tools — it saves decisions with remember and pulls relevant history with recall, with no user action:
// Claude Code mcp config
{
"mcpServers": {
"dolores": {
"command": "node",
"args": ["/abs/path/to/dolores/packages/mcp/dist/index.js"],
"env": {
"DOLORES_WORKSPACE_ID": "00000000-0000-0000-0000-000000000001",
"DOLORES_DAEMON_PORT": "4505"
}
}
}
}
Now memory is not a passive store — it's a tool the agent actively uses across sessions.
How it works
Structural (facts) |
Semantic (memories) |
|
|---|---|---|
| Stored as | key/value | free text + 384d embedding + tsvector |
| Retrieved by | exact SQL, no embedding | pgvector cosine + full-text (RRF) |
| Conflict resolution | ON CONFLICT upsert (last writer wins) |
supersede on >0.9 cosine similarity |
| Example | stack/db = Postgres + pgvector |
"migration ordering bug fixed last deploy" |
Isolation. Every row carries workspace_id (+ optional user_id) and is guarded by Postgres Row-Level Security. The daemon connects as a non-superuser and sets the tenant per transaction, so cross-tenant reads are impossible — not just discouraged.
Decay. A daily pg_cron job softens the importance of stale, un-recalled memories — like human memory. Deletion is off by default (DOLORES_DECAY_MODE=conservative); the aggressive delete policy is opt-in.
Where it fits
Mem0, Letta (ex-MemGPT), and Zep all do Postgres/vector agent memory — the concept is proven. dolores's niche: self-hosted, free local embeddings, MCP-native, no raw-transcript storage, KVKK/GDPR-clean. Your subscription, your database, nothing else.
Tech stack
TypeScript (strict) · Node ESM · pnpm monorepo · Prisma + raw SQL for pgvector · fastembed · fastify · commander · @modelcontextprotocol/sdk · zod · Docker Compose.
| Package | Responsibility |
|---|---|
@dolores/db |
Prisma schema, raw-SQL migration (pgvector + pg_cron), Dockerfile, RLS, withTenant |
@dolores/core |
embedder abstraction · hybrid retrieval · extraction (owns the shared contracts) |
@dolores/daemon |
loads the embedder once, owns the pool, serves localhost HTTP |
@dolores/cli |
commander-based thin client |
@dolores/mcp |
MCP server (remember / recall) |
Examples
Runnable integration examples (MCP wiring, context injection, Node.js SDK): examples/
Development
pnpm install
pnpm build # build every package
pnpm test # run every package's tests
pnpm lint # biome
pnpm db:up # start Postgres locally
Architecture decisions and their why live in MEMORY.md; repo working rules are in CLAUDE.md.
→ Backup / restore, decay modes, daemon env vars, and production hardening: docs/OPERATIONS.md.
Contributing
Issues and PRs are welcome. Keep the architectural invariants intact: no raw transcripts, embeddings behind the Embedder interface, the LLM off the critical path, and Postgres as the single source of truth. See CLAUDE.md.
License
MIT © 2026 Yener Yiğit Çelik
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.