fleet-mem

Shared code intelligence for agent fleets. AST-aware semantic search, multi-agent memory, and git-concurrent coordination.

When multiple AI agents work on the same codebase, they fight. Agent A rewrites a function that Agent B is also modifying. Agent C searches for a pattern that Agent D already found and documented. Agents repeat work, create conflicts, and operate on stale information.

fleet-mem is a local MCP server that gives AI coding agents shared context:

• Zero data leakage by default. Runs entirely on your machine using local Ollama embeddings. No cloud APIs, no telemetry, no data leaves your network. Cloud embedding providers (OpenAI, Gemini, etc.) are available as an opt-in choice.
• Token-efficient code search. Understands the structure of your code via Abstract Syntax Trees (AST). Returns the specific function, not the entire file.
• Shared memory across agents. Agent A discovers "this service uses JWT, not sessions." Agent B finds that knowledge automatically when working on the same code. Memories persist across sessions.
• Fleet-aware coordination. Agents declare what files they are working on, get blocked on conflicts before they start, and get notified when another agent's merge affects their context.

---

Getting started

Prerequisites

• Python 3.11+
• Ollama running locally (brew, systemd, or Docker)
• ollama pull nomic-embed-text

Install

pip install fleet-mem

Or from source:

git clone https://github.com/sam-ent/fleet-mem.git
cd fleet-mem
./scripts/setup.sh  # Creates venv, installs deps, registers MCP server

Docker (alternative)

./scripts/docker-setup.sh

MCP client configuration for Docker:

{
  "mcpServers": {
    "fleet-mem": {
      "command": "docker",
      "args": ["exec", "-i", "fleet-mem-fleet-mem-1", "python", "-m", "fleet_mem.server"]
    }
  }
}

Mount your code as a volume to index it:

# Add to docker-compose.yml under fleet-mem.volumes:
- /path/to/your/projects:/projects:ro

Index your codebases

./scripts/index-repos.sh --root ~/projects

MCP client configuration

Add to your MCP client settings (the setup.sh script does this automatically for the default client):

{
  "mcpServers": {
    "fleet-mem": {
      "command": "python3",
      "args": ["-m", "fleet_mem.server"],
      "env": {
        "OLLAMA_HOST": "http://localhost:11434",
        "ANONYMIZED_TELEMETRY": "False",
        "FLEET_STATS_SOCK": "~/.fleet-mem/stats.sock"
      }
    }
  }
}

Important: Do not set "cwd" in the MCP config. Claude Code sets the working directory to the project the agent is working in. A hardcoded cwd breaks worktree detection and auto-coordination.

fleet-mem works with any MCP-compatible client. Your client starts it automatically on the first tool call.

Example agent queries

Once indexed, agents can ask things they could not do with grep:

• "Find the authentication middleware and show me how tokens are validated"
• "Which agent is currently working on the database schema?"
• "What did other agents learn about the payment gateway this session?"
• "If I merge this branch, which agents will have stale context?"

---

How it works

fleet-mem installs once as a global MCP server. It can index any number of projects. Each project gets its own collection in ChromaDB. All agents share the same server instance.

~/projects/
  project-a/     indexed as code_project-a
  project-b/     indexed as code_project-b
  project-c/     indexed as code_project-c

~/.local/share/fleet-mem/
  chroma/              vector embeddings (shared)
  memory.db            agent memories + FTS + file anchors (WAL mode)
  fleet.db             locks, subscriptions, sessions, notifications (WAL mode)
  embeddings_cache.db  embedding vector cache

Architecture

Components

Component	What it is	Why we chose it
Ollama	Local ML inference server	Runs embedding models on your machine at zero cost. Supports dozens of models. Works via Docker, systemd, or brew. Swappable via the Embedding base class
ChromaDB	Vector database (HNSW)	Purpose-built for similarity search over embeddings. Runs in-process, no separate server needed
SQLite + FTS5	Relational database with full-text search	Agent memories need both keyword search and structured queries. FTS5 + ChromaDB vectors give hybrid ranking via reciprocal rank fusion
tree-sitter	Incremental parsing library	Splits code into semantic chunks (functions, classes, methods) instead of arbitrary character windows. Search results are meaningful code units, not fragments
xxHash (xxh3_64)	File change detection + chunk IDs	Detects which files changed between sync cycles. Not a security function, purely for diffing. ~10x faster than SHA-1

Language support

Language	Splitting method	Support level
Python, TypeScript, JavaScript	AST-aware	Tier 1: functions, classes, methods
Go	AST-aware	Tier 2: functions, methods, types
Rust	AST-aware	Tier 2: functions, impl blocks, structs, enums, traits
All other languages	Text-only	Fallback: sliding window (2500 chars, 300 overlap)

AST-aware splitting means search results are complete, meaningful code units. Text-only fallback still works but may return partial functions. Adding a new language requires defining its tree-sitter node types in fleet_mem/splitter/ast_splitter.py (contributions welcome).

Process flows

Indexing a codebase

Problem: Agents read entire files to understand code, burning tokens and missing context across files.

Solution: One-time indexing parses code into semantic chunks and embeds them. Agents search by meaning across the whole codebase.

Semantic code search

Problem: Grep requires exact strings. Agents don't know file names or function signatures in unfamiliar code.

Solution: Natural language query returns ranked code snippets with file paths and line numbers. No exact match needed.

Storing and searching memory

Problem: Agents lose everything they learn when a session ends. The next agent re-discovers the same things from scratch.

Solution: Discoveries persist in a shared memory store. Any agent can find them later via keyword or semantic search.

File locking

Problem: Concurrent agents modify the same files, causing merge conflicts and wasted work.

Solution: Agents automatically lock files they've modified on their branch. Conflicts are caught immediately, not after hours of wasted effort. If some files overlap with another agent's lock, only the non-conflicting files are locked.

Cross-agent knowledge sharing

Problem: Agent A discovers something important about the code. Agent B, working in the same area, has no way to know.

Solution: Agents subscribe to file patterns they care about. When another agent stores a discovery matching that pattern, subscribers are notified automatically.

Merge impact preview

Problem: Agent A merges a PR. Agents B and C are still working on branches that now have stale context. No one tells them.

Solution: Before merging, see exactly which agents, memories, and branches will be affected. After merging, one call notifies everyone and marks stale context.

Embedding providers

The default is Ollama (local, free). fleet-mem also ships an OpenAI-compatible adapter that works with any provider offering an OpenAI-style embeddings API.

Provider	Setup	Cost
Ollama (default)	Install Ollama, ollama pull nomic-embed-text	Free
OpenAI	Set EMBEDDING_PROVIDER=openai-compat, EMBED_API_KEY, EMBED_MODEL=text-embedding-3-small	~$0.02/1M tokens
DeepSeek	Set EMBED_BASE_URL=https://api.deepseek.com/v1, EMBED_API_KEY, EMBED_MODEL=deepseek-embed	~$0.01/1M tokens
Gemini	Set EMBED_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/, EMBED_API_KEY, EMBED_MODEL=text-embedding-004	Free tier available
Together	Set EMBED_BASE_URL=https://api.together.xyz/v1, EMBED_API_KEY, model of choice	Varies
Local vLLM	Set EMBED_BASE_URL=http://localhost:8000/v1, no API key needed	Free

See .env.example for full configuration. For providers without an OpenAI-compatible API (Cohere, AWS Bedrock, Hugging Face), see docs/custom-embedding-providers.md. The adapter interface is four methods and typically under 30 lines.

---

Features

Code understanding

• Semantic search: "find auth middleware" returns relevant functions, not string matches
• Symbol lookup: find function/class definitions across indexed projects
• Dependency analysis: trace what calls or imports a given symbol
• Incremental sync: xxHash Merkle tree detects file changes, re-indexes only deltas
• Branch-aware indexing: overlay collections for feature branches keep changes isolated from the main index

Fleet coordination

• Auto-coordination: on startup and every 30s, agents automatically lock and subscribe to files modified on their branch vs main. No manual lock_acquire or memory_subscribe calls needed
• Per-file lock filtering: if some files overlap with another agent's lock, only the non-conflicting files are locked (not all-or-nothing)
• Worktree-aware: agents in different git worktrees of the same repo share a single project namespace. Uses git rev-parse --git-common-dir with --show-toplevel fallback
• UPSERT locks: one lock per agent per project, atomically updated via INSERT...ON CONFLICT DO UPDATE. No duplicate accumulation across restarts
• Cross-agent memory: agents share discoveries via subscriptions and notifications. memory_store automatically notifies subscribers matching the file path
• Merge impact preview: before merging, see which in-flight agents would be affected
• Post-merge notification: after merging, automatically notify affected agents, mark stale file anchors, and release locks on the merged branch
• WAL mode + busy timeout: all fleet DB modules use SQLite WAL mode with 5s busy timeout for concurrent multi-agent access

---

Configuration

All settings via environment variables or a .env file in the project root. Copy .env.example to get started.

Variable	Default	Description
OLLAMA_HOST	http://localhost:11434	Ollama API endpoint
OLLAMA_EMBED_MODEL	nomic-embed-text	Embedding model name. Switching this between models with different output dimensions requires dropping affected collections — see Switching embed models below.
EMBEDDING_PROVIDER	ollama	Provider: ollama or openai-compat
CHROMA_PATH	~/.local/share/fleet-mem/chroma	ChromaDB storage
MEMORY_DB_PATH	~/.local/share/fleet-mem/memory.db	Agent memory database
FLEET_DB_PATH	~/.local/share/fleet-mem/fleet.db	Fleet coordination database
SYNC_INTERVAL	300	Background code index sync (seconds)
FILE_WATCHING	true	Enable filesystem watching for near-instant sync
FLEET_MEM_MAX_CHUNK_CHARS	5000	Max characters per embed chunk. Oversized chunks are recursively subdivided before hitting the embed model. Default (~1250 tokens for English) fits comfortably below a 2048-token context window. Lower this if using non-English content or a smaller-context embed model.
FLEET_MEM_MAX_CHUNK_TOKENS	unset	Optional, opt-in (issue #42). Max tokens per embed chunk, measured with the embed model's actual tokenizer. When set, the chunker subdivides chunks until each fits within this token cap in addition to the char cap. Recommended: ~80% of the model's max position embeddings (e.g. 400 for all-minilm (512 tok), 1600 for nomic-embed-text (2048 tok)). Requires the tokenizer-aware extra: pip install fleet-mem[tokenizer-aware]. Falls back silently to the char cap if the tokenizers package is missing, the model is not in the built-in mapping (all-minilm, nomic-embed-text, mxbai-embed-large, bge-large, bge-m3, snowflake-arctic-embed), or the HF tokenizer fails to load. Use this to eliminate the O(log)-per-chunk bisect+mean-vector recovery overhead on dense content (code, non-English text, base64-like blobs) where char-cap alone underestimates token count.

Switching embed models

ChromaDB collections are dim-locked at creation — the vector dimension is fixed when a collection is first created and cannot change. If you switch OLLAMA_EMBED_MODEL (or EMBED_MODEL for cloud providers) to a model with a different output dimension (e.g. all-minilm 384-dim ↔ nomic-embed-text 768-dim), fleet-mem detects the mismatch on the affected collection's next insert or query and raises fleet_mem.vectordb.errors.DimMismatchError (added in #47).

The exception carries model_name, model_dim, collection_name, and collection_dim for programmatic recovery (e.g. an automated re-indexer that catches the error and decides whether to drop + rebuild).

To switch models cleanly, drop the affected collections and re-index:

# Drop a single project's index via the MCP tool
clear_index(path="/path/to/project")

Or, for a wholesale reset, remove the on-disk ChromaDB directory and let fleet-mem re-create collections at the new dim on next index:

rm -rf ~/.local/share/fleet-mem/chroma/

Without dim-mismatch detection the failure mode would be either a late ChromaDB error during inserts (recoverable but wasteful) or — worse — silent garbage retrieval at query time, since cosine distance against mismatched-dim vectors is meaningless. DimMismatchError exists to fail fast and steer you to a clean recovery path.

Background sync timing

What	Timing	How
Code index refresh	Every SYNC_INTERVAL seconds (default: 300)	Polls filesystem, computes xxHash digests, re-indexes changed files
Agent memory writes	Immediate	Direct SQLite + ChromaDB insert on memory_store call
Lock acquire/release	Immediate	Direct SQLite UPSERT (one lock per agent per project)
Auto-coordination	Every 30s (heartbeat)	Detects new commits via git diff main...HEAD, locks + subscribes to changed files
Notifications	Immediate	Created on memory_store if subscriptions match (scoped by project)
Session heartbeat	Every 30s	Updates last_activity_at, extends lock TTLs
Stale session pruning	On query	Sessions disconnected >24h are auto-deleted

For fast-moving multi-agent work, reduce SYNC_INTERVAL to 30-60. File-watching is also available for near-instant sync — set FILE_WATCHING=true (the default) to detect changes immediately without polling.

Scripts

Script	Purpose
scripts/setup.sh	One-time install: venv, dependencies, Ollama check, MCP registration
scripts/index-repos.sh	Find git repos under a root directory and index each one
scripts/import-flat-files.py	Import existing memory files (markdown with YAML frontmatter)
scripts/embed-existing-nodes.py	Embed existing memory DB nodes into ChromaDB for semantic search

scripts/index-repos.sh

Iterates git repositories under --root (max depth 2) and indexes each one. Per-repo failures don't halt the loop by default — the script prints FAILED to index <name>. Continuing... and moves on to the next repo. After all repos are attempted, it prints an Indexed N/M repos. summary and exits non-zero if any repo failed (since #33).

For CI / orchestration use cases that want to abort on the first failure, set FAIL_FAST=1:

FAIL_FAST=1 ./scripts/index-repos.sh --root ~/projects

Exit codes:

• 0 — all repos indexed successfully
• 1 — one or more repos failed (default mode finished attempting every repo)
• 2 — FAIL_FAST=1 triggered an early exit on the first failure

---

Troubleshooting

DimMismatchError raised after switching embed models

You changed OLLAMA_EMBED_MODEL (or EMBED_MODEL) to a model with a different output dimension. ChromaDB collections are dim-locked at creation. See Switching embed models for the recovery path.

Indexer logs ollama: context-length 400 on batch of N; bisecting (depth=...)

Normal recovery behavior. The chunker emitted a chunk that exceeds the embed model's token-context window — the embedder bisects the batch (and, if needed, splits the offending text and averages the resulting vectors) until each input fits. Indexing continues and no data is lost.

If you see this frequently and want to eliminate the recovery overhead, switch to tokenizer-aware mode by setting FLEET_MEM_MAX_CHUNK_TOKENS (see Configuration). This trades a small tokenizers dependency for elimination of the bisect+mean-vector recovery path on dense content (code, non-English text, base64-like blobs).

Already indexed (N chunks). Use --force to re-index. on retry

The scripts/index-repos.sh short-circuit: if a code_<project> collection already exists with chunks in it, the script skips re-indexing that repo. This is the correct behavior for incremental runs — fleet-mem's merkle-tracked sync (every SYNC_INTERVAL seconds) handles deltas separately. To force a full re-index of a specific project, drop its collection first:

# Via the MCP tool
clear_index(path="/path/to/project")

Then re-run scripts/index-repos.sh.

Indexer hangs without progress output

The progress callback fires every 50 files. Most likely the chunker is processing a single very large file or a deeply-nested directory. If the run pauses between callbacks for more than a few minutes, attach py-spy dump --pid <indexer-pid> to inspect the Python stack. If reproducible, please open an issue with a minimal reproducer.

---

Observability

fleet-mem includes OpenTelemetry tracing, structured logging with trace correlation, and a terminal monitoring UI. All disabled by default.

Observability architecture

Monitoring a fleet

Problem: Multiple agents are working on the same codebase. You can't tell which agents are active, what files they've locked, or whether they're conflicting — until something breaks.

Solution: Agents register on connect. The TUI monitor polls fleet state over a Unix socket and shows agents, locks, subscriptions, and notifications in real time.

Quick start — Fleet monitor TUI

No external infrastructure needed. Install the monitor extra, set one env var, and go:

# 1. Install with monitor
pip install fleet-mem[monitor]

# 2. Enable the stats socket (add to your .env or MCP server config)
FLEET_STATS_SOCK=~/.fleet-mem/stats.sock

# 3. Launch the monitor in a separate terminal
fleet-mem monitor

The TUI connects via a Unix domain socket (0600 permissions — only the socket owner can connect, no network exposure). It shows:

• Dashboard tab: Aggregate metrics with sparklines, conflict alerts (red banner when agents have overlapping locks), and gauges for active agents/locks/notifications/memory
• Agents tab: Registered agents with project, worktree, branch, and color-coded rows (green=active+locked, white=active, dim=disconnected). Disconnected agents hidden by default — press d to toggle
• Locks tab: Active file locks with per-file conflict visualization showing which files are contested between which agents
• Subscriptions tab: Tree view grouped by agent (collapsible — agent-alpha (30 files) → expand), not a flat table of 800+ rows
• Memory tab: Memory node counts, file anchors, embedding cache, and collection stats
• Notifications tab: Cross-agent notifications with read/unread styling
• Log tab: Live event log showing auto-coordination events, prune actions, and toggle state changes

Keybindings: q quit, r refresh, f filter by agent ID, d toggle disconnected agents, x prune disconnected sessions from DB

For Docker deployments, the socket is exposed via a named volume (fleet-sock). Mount it on the host to run the monitor:

docker compose up -d
fleet-mem monitor --sock /var/lib/docker/volumes/fleet-mem_fleet-sock/_data/stats.sock

Advanced — OpenTelemetry tracing

For teams with existing observability infrastructure (Jaeger, Grafana Tempo, Datadog), fleet-mem exports OpenTelemetry spans:

OTEL_ENABLED=true
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317  # any OTLP-compatible collector

Data plane spans:

Span	Key attributes
fleet.index	project, chunk_count
fleet.search	query_hash (never raw query), result_count, cache_hits
fleet.memory.store	content_hash, node_type
fleet.memory.search	query_hash, result_count

Coordination plane spans:

Span	Key attributes
fleet.lock.acquire	agent_id, project, conflict_count, lock_id
fleet.lock.release	agent_id, project, released_count
fleet.lock.query	project, lock_count
fleet.lock.heartbeat	agent_id, extended_count
fleet.memory.feed	agent_id, since_minutes, result_count
fleet.memory.subscribe	agent_id, project, subscription_count
fleet.memory.notifications	agent_id, notification_count
fleet.memory.notify_subscribers	author_agent_id, subscriber_count, notification_count
fleet.merge.impact	project, file_count, conflict_count, subscriber_count
fleet.merge.notify	project, branch, notification_count, stale_anchor_count

All content is hashed in span attributes for privacy. Raw code and queries never appear in traces.

Structured logging

fleet-mem uses structlog with OpenTelemetry trace context injection. When a span is active, trace_id and span_id are automatically added to every log line — enabling log-to-trace correlation in Grafana, Datadog, or any log aggregator.

• OTEL_ENABLED=true: JSON output (machine-parseable, for log pipelines)
• OTEL_ENABLED=false (default): Human-readable console output

Fleet stats (no collector needed)

The fleet_stats MCP tool returns current metrics without requiring an external collector:

fleet_stats() -> {
  collections: {code_myproject: 1523},
  total_chunks: 1523,
  memory_nodes: 47,
  active_locks: 2,
  subscriptions: 5,
  pending_notifications: 1,
  cached_embeddings: 892
}

---

MCP tools reference

Code search (6 tools)

Tool	Parameters	Description
index_codebase	path, branch?, force?	Index a codebase (background). Branch-aware when branch specified
search_code	query, path?, branch?, limit?	Semantic code search across indexed projects
find_symbol	name, file_path?, symbol_type?	Find symbol definitions (functions, classes)
find_similar_code	code_snippet, limit?	Find code similar to a given snippet
get_change_impact	file_paths?, symbol_names?	Find code affected by changes to given files/symbols
get_dependents	symbol_name, depth?	Trace what calls/imports a symbol (BFS)

Agent memory (4 tools)

Tool	Parameters	Description
memory_store	node_type, content, summary?, keywords?, file_path?, line_range?, source?, project_path?	Store a memory with optional file anchor
memory_search	query, top_k?, node_type?	Hybrid keyword + semantic memory search
memory_promote	memory_id, target_scope?	Promote a project memory to global scope
stale_check	project_path?	Find memories whose anchored files have changed

Fleet coordination (10 tools)

Tool	Parameters	Description
fleet_register	agent_id, project, worktree_path?, branch?	Register an agent session (call once when starting work)
fleet_agents		List all registered agents with status (active/idle/disconnected)
lock_acquire	agent_id, project, file_patterns	Declare files an agent is working on
lock_release	agent_id, project	Release file locks
lock_query	project, file_path?	Check who holds locks on which files
merge_impact	project, files	Preview which agents/memories are affected by a merge
notify_merge	project, branch, merged_files	Post-merge: notify affected agents, mark stale anchors
memory_feed	agent_id?, since_minutes?	Recent memories from other agents
memory_subscribe	agent_id, file_patterns	Subscribe to memories about specific files
memory_notifications	agent_id	Check for new relevant memories from other agents

Status and observability (7 tools)

Tool	Parameters	Description
get_index_status	path	Check indexing status for a project
clear_index	path	Drop a project's index and reset
get_branches	path	List indexed branches with chunk counts
cleanup_branch	path, branch	Drop a branch overlay after merge
fleet_stats		Current metrics: chunks, memories, locks, cache hits, notifications
reconcile	path	Remove ghost chunks whose source files no longer exist
clear_embedding_cache		Clear the embedding vector cache, forcing re-embedding on next use

---

What's next

• Go/Rust recursive AST splitting (promote to Tier 1)
• Performance benchmarks on real codebases
• MCP client configuration guides for Cursor, Windsurf
• OTel Metrics API (histograms/counters for coordination)
• Grafana dashboard JSON for coordination observability

See roadmap.md for the full plan.

---

License

AGPL-3.0

Acknowledgments

Architecture inspired by claude-context by Zilliz (MIT License). Design patterns informed by their TypeScript reference (vector database abstraction, embedding adapter, Merkle DAG, AST splitter). All code is an original Python implementation with significant additions (agent memory, fleet coordination, hybrid search, staleness detection).