SafeFlo

CI License: MIT Node.js No install scripts

A local, transparent agent memory server for Claude Code. MCP server with persistent hybrid memory (lexical + semantic search), a memory lifecycle (types, importance, supersession, consolidation), task planning, and logical agent coordination — all data in ./.safeflow/, no hidden install scripts, no modification of files outside the project.

🇷🇺 Read in Russian: README.ru.md

TL;DR

Give Claude Code persistent, searchable memory that understands paraphrases — plus structured plans and logical agents. All data lives in ./.safeflow/. Uninstall with a single command. No surprises.

Contents

• What's inside
• Installation
• Using from Claude Code
• Memory model
• Search
• Benchmark
• Programmatic API
• Trade-offs and limitations
• Uninstall
• Security
• Contributing
• License

What's inside

Module	Description
Memory store	SQLite with hybrid search — FTS5 (lexical) + sqlite-vec (semantic) fused via Reciprocal Rank Fusion. Memory lifecycle: types, importance, supersession, consolidation. Parameterized SQL only.
Task planner	Structured goal decomposition into steps with dependencies and verified status transitions.
Agent coordinator	Registration of logical agents with an isolated memory namespace each. No background processes.
MCP server	18 tools with transparent, functional-only descriptions.
Audit log	Append-only JSONL of all operations.
CLI	init, status, mcp, backfill-embeddings, uninstall (with real, complete cleanup).

Installation

# Clone the repository — no curl | bash installers.
git clone https://github.com/G1ngercy/SafeFlo.git
cd safeflow

# npm ci strictly follows package-lock.json — no version substitution.
# package.json contains no preinstall/postinstall scripts.
npm ci

# Build and test
npm run build
npm test

To use in your project:

cd /path/to/your/project
node /path/to/safeflow/dist/cli.js init

This creates:

• ./.safeflow/ — local databases, model cache, and audit log
• ./.claude/commands/safeflow-*.md — slash commands for Claude Code

To register the MCP server with Claude Code:

claude mcp add safeflow -- node /path/to/safeflow/dist/mcp/server.js

Using from Claude Code

Once connected, Claude Code gains these tools:

Memory:

• memory_store(namespace, key, content, metadata?, memory_type?, importance?, source?)
• memory_get(namespace, key)
• memory_recall(namespace, query, limit?, memory_types?, include_superseded?) — hybrid search (recommended)
• memory_search(namespace, query, limit?) — [deprecated, use memory_recall] FTS5-only
• memory_list(namespace, limit?)
• memory_delete(namespace, key)
• memory_supersede(old_id, new_content, reason) — replace an outdated fact
• memory_consolidate(namespace, dry_run?) — find episodic clusters to summarize

Planning:

• plan_create(goal)
• plan_add_step(planId, title, description, dependsOn?)
• plan_update_step_status(stepId, status)
• plan_get(planId)
• plan_ready_steps(planId) — steps that are ready to start
• plan_list(limit?)

Agents:

• agent_register(role, task?)
• agent_list(status?)
• agent_update_status(agentId, status)

Audit:

• audit_tail(n?) — last N events from the audit log

And slash commands: /safeflow-plan, /safeflow-memory, /safeflow-agents.

Memory model

Episodic / semantic / procedural. Every record has a memory_type. Episodic is the default: a specific observation or event ("we decided X today"). Semantic is generalized, durable knowledge distilled from episodes. Procedural captures how to do something — steps, conventions, runbooks. The type nudges ranking and is the unit consolidation promotes (episodic → semantic).

Importance. Each record carries an importance in [0, 1] that boosts ranking in recall. It defaults to a transparent, content-derived heuristic (type, decision keywords in RU/EN, length) and can be set explicitly. No machine learning, no hidden signals.

Supersession. Facts go stale. memory_supersede(old_id, new_content, reason) writes the replacement as a new record and marks the old one superseded_by the new one. The old record is kept for history and audit but excluded from recall by default (pass include_superseded to see it).

Consolidation. memory_consolidate finds clusters of similar, older episodic records (greedy agglomeration over their vectors by cosine similarity) and returns them with sample contents. The server performs no summarization and no network calls — the client decides what to summarize and stores the result as a semantic record. This keeps the "no network at runtime" boundary intact.

Search

Hybrid: FTS5 for lexical matching + sqlite-vec for semantic similarity, combined via Reciprocal Rank Fusion (RRF), then adjusted by importance and a recency boost. If the embedding model is not present (or you opt out), recall degrades gracefully to FTS5-only — nothing breaks, you just lose the semantic leg.

Benchmark

npm run bench compares the legacy FTS-only search() against the v2 hybrid recall() over a mixed RU/EN dataset (6 cases, 24 records, 22 queries). Results with paraphrase-multilingual-MiniLM-L12-v2:

Query type	v1 recall@5	v2 recall@5	Δ
lexical	100.0%	100.0%	+0.0 п.п.
synonym	44.4%	100.0%	+55.6 п.п.
concept	57.1%	100.0%	+42.9 п.п.

MRR on synonym queries rises from 0.333 to 0.861. As expected, lexical queries are unchanged (FTS already nails exact words); the win is on paraphrased and conceptual queries — exactly where a key-value/FTS store falls short for an AI agent. Raw results are in benchmark-results/.

Programmatic API

import {
  MemoryStore,
  TaskPlanner,
  AgentCoordinator,
  AuditLogger,
} from "safeflow";

const audit = new AuditLogger(process.cwd());
const memory = new MemoryStore(process.cwd(), audit);
const planner = new TaskPlanner(process.cwd(), audit);
const coord = new AgentCoordinator(process.cwd(), audit);

await memory.store("project.notes", "decision-1", "Use SQLite for memory", {}, {
  memoryType: "semantic",
});

const hits = await memory.recall({
  namespace: "project.notes",
  query: "which database did we choose",
});

const plan = planner.createPlan("Add authentication");
const step = planner.addStep(plan.id, {
  title: "Design schema",
  description: "users, sessions",
  dependsOn: [],
});

const agent = coord.register("researcher", "Survey auth libraries");

Trade-offs and limitations

• First model load requires the network. Semantic search relies on the paraphrase-multilingual-MiniLM-L12-v2 model (~120MB), downloaded once from Hugging Face into ./.safeflow/models/. Until then, search is FTS5-only. You can opt out of embeddings entirely and stay FTS-only. See SECURITY.md.
• Native modules. better-sqlite3 and sqlite-vec are native; they need prebuilt binaries or a toolchain for your platform.
• Scale. This is a local, single-file SQLite design. Past ~100k records you want a dedicated vector database and a different architecture; SafeFlo is built for a project's working memory, not a data lake.

Uninstall

node /path/to/safeflow/dist/cli.js uninstall --yes

This removes:

• ./.safeflow/ — all local databases, model cache, and audit log
• ./.claude/commands/safeflow-*.md

SafeFlo uses no global paths whatsoever, so there is nothing to clean up outside the project. Genuinely nothing. Verify for yourself: grep -rn "homedir\|os\.home" src/ returns no results.

Security

The full threat model and guarantees are in SECURITY.md. Quick summary:

• No install scripts in package.json (CI checks this automatically).
• No network calls during memory operations. One-time exception: the embedding model (~120MB) is downloaded on first use; opt out to stay FTS5-only. Documented in SECURITY.md.
• No modification of files outside the project — all data in ./.safeflow/.
• Parameterized SQL everywhere, Zod validation on every input.
• Protection against path traversal (including ....// bypasses), prototype pollution, SQL injection.
• Transparent MCP descriptions — no hidden directives to the LLM. Audited automatically in CI.
• Idempotent migrations with automatic backups to ./.safeflow/backups/ before any schema change.
• Complete uninstall with a single command.
• Pinned dependencies — 5 packages with exact versions (@modelcontextprotocol/sdk, better-sqlite3, sqlite-vec, @xenova/transformers, zod).
• Distribution via git clone — no npm registry, no curl | bash installer. You build from a pinned package-lock.json; releases are tagged commits with GitHub Release notes (see docs/RELEASE.md).

Vulnerabilities — through private security advisory, not through public issues. See SECURITY.md.

Contributing

See CONTRIBUTING.md. In short:

• For bugs — issue → fork → PR with a test.
• For features — issue first, then PR.
• For vulnerabilities — private security advisory, not a public issue.

Code of Conduct — CODE_OF_CONDUCT.md.

License

MIT, see LICENSE.