How BrigadeStack compares
Pick the question you came with. Each answer says when another tool is the better fit. Below them, a feature matrix shows what we checked in each neighbor's own docs.
Start with your question
-
How do I keep memory across agents?
Durable context that survives sessions and moves between agents, and how that differs from a memory service.
- Memory handoffs across coding agents One handoff format, one ingest path
- Stop Claude Code memory and bootstrap bloat Slim bootstrap files, cards for the rest
- BrigadeStack vs mem0 A memory service or reviewed local files
- BrigadeStack vs agentmemory Automatic capture or reviewed handoffs
-
How do agents see code structure?
Callers, callees, and blast radius for agents, built from your repo on your machine.
- Local code graph for AI coding agents Impact queries without a hosted service
-
How do I prove what an agent did?
Searchable evidence of agent work, and skill scores that follow the text a skill has now.
- Local evidence ledger for agent work Search sessions and receipts locally
- Keep an agent skill's score honest after you edit it Scores that reset when the text changes
-
Which tool fits, and which Brigade is this?
A field guide to the neighbors by job, and how BrigadeStack differs from the other projects named Brigade.
- Best tools for AI agent config, MCP, and memory The neighbors, grouped by job
- Which Brigade is this? BrigadeStack, not the Kubernetes one
At a glance
BrigadeStack's cells for each job, and where a neighbor leads. The matrices below hold every cell and its wording.
| Job | BrigadeStack, one square per column | Where a neighbor leads |
|---|---|---|
| Work and orchestration 5 neighbors checked | | Beads for a distributed Dolt-backed task store. Gas Town for a multi-agent runtime with scheduler caps and a merge queue. ActiveGraph for replay. |
| Memory 6 neighbors checked | | Mem0 for hosted recall and Dream consolidation. agentmemory for automatic hook capture. Hindsight and Letta Code for reflective learning. |
| Tools and configuration 5 neighbors checked | | Rulesync and agentsync for projecting rules, skills, hooks, and subagents. Ruler for rules, skills, and experimental subagents. add-mcp for a quick install across 25 clients. |
| Evidence and operations 3 neighbors checked | | ActiveGraph for replay, fork, and diff. Graphiti for the validity window of a fact. |
Feature matrix
- Supported
- The checked official sources document the capability in the scope the cell names.
- Partial
- The sources document a subset, an indirect route, or something the operator or application has to build.
- Not documented
- Not found in the checked sources. This is not proof the project lacks it.
Work and orchestration
| Project | Ready work | Claim safety | Parallel and run control | Task-store boundary | Verification |
|---|---|---|---|---|---|
| BrigadeStack | Supported Dependency-gated ready set + parallel-safe waves | Supported Fail-closed local CAS + optional generation-fenced hub claims | Supported Wave dispatch + --watch, worker caps, deadline/launch ceilings, retry breaker | Partial Per-repo JSON ledger + Beads import. Opt-in Hub Worklore DB, deps stored only | Supported Command + exit code + Git receipts 0.32.0 beta Adds observe mode: hooks record the checks agents run, with a first-pass report by harness and model |
| Beads | Supported Dependency-unblocked tasks | Supported Atomic claim | Partial Multi-agent workflows + Dolt sync | Supported Dolt-backed distributed store | Partial Task history + audit trail |
| Gas Town | Supported Beads-backed ready work | Supported Local dispatch + assignee locks | Supported Scheduler caps, pause/resume + agents | Supported Beads-backed task store | Supported Refinery verification + merge gates |
| nWave | Partial S/M/L assessment + next-step guide | Not documented | Partial Per-feature worktree lanes, no cap | Partial Per-feature handoff.json in repo | Partial Oracle, RED/GREEN + review, run by the host |
| ActiveGraph | Supported Reactive graph triggers | Partial Optimistic patches, app-defined claims | Supported Bounded runs + cooperative drains | Supported Append-only event log | Partial Event trace + replay |
| Claude Code agent teams | Supported Shared task list + dependencies | Supported File-locked claims | Supported Teammates + optional plan mode (experimental) | Partial Local, one team per session | Partial Hook gates, checks user-built |
Memory
| Project | Durable memory | Capture and recall | Care and operations | Learning | Provenance |
|---|---|---|---|---|---|
| BrigadeStack | Supported Shared files + cards, optional hub | Supported Linted handoffs + bounded recall | Supported Owner policy, care review queue, vault project/search/propose 0.32.0 beta Memory care tends cards on its own: date fixes and review inbox drain, journaled with undo | Partial Manifest-scored promote + rollback, reflection proposals, gated auto-apply off by default | Supported Work receipts + sources |
| Mem0 | Supported User + agent + session stores | Supported Fact extraction + hybrid search | Supported Edit, delete, history, expiry + export | Supported Dream: merge, supersede + pattern synthesis | Partial IDs, timestamps + change history |
| Letta Code | Supported Git-backed memory filesystem | Supported Agent-edited memory files + message search | Supported Doctor audit, backup before reorganize | Supported Dreaming + skill learning | Partial Git commit per edit, authorship not documented |
| agentmemory | Supported Shared coding-agent memory | Supported Hooks + hybrid and graph recall | Supported Versioning, retention + rollback | Supported Reflection, lessons + consolidation | Supported Origin channels + source citations |
| Hindsight | Supported Facts, experiences + mental models | Supported Retain, recall + reflect | Supported Document history + model refresh | Supported Reflective memory learning | Supported Document/chunk citations + metadata |
| Built-in harness memory | Supported Tool-specific files + stores | Partial Rules + opt-in or automatic recall | Partial Per-tool inspection + controls | Partial Saved corrections + consolidation | Partial Files + prior-chat evidence, varies by tool |
| OpenClaw | Supported Plain files + index | Supported Hybrid search + optional session recall | Supported Consolidation, review + forgetting | Supported Dreaming + gated promotion | Supported Origin metadata + change history |
Tools and configuration
| Project | MCP | Reviewed rules and skills | Projection behavior | Source of truth | Audit and rollback |
|---|---|---|---|---|---|
| BrigadeStack | Supported MCP + tools catalog | Supported Skills + adapters | Supported Dry-run first, field-fidelity report, then apply | Supported BrigadeStack catalog | Supported Preserve + gate + recoverable projections |
| add-mcp | Supported MCP servers across 25 native client configs | Not documented | Supported Native field mapping + loss warnings | Partial Registry + native files | Partial List/remove/sync + dropped-field warnings |
| Rulesync | Supported MCP per target | Supported Rules, skills, hooks + commands | Supported Per-target generation + dry-run | Supported .rulesync source tree | Partial Dry-run, --check + cleanup |
| chezmoi | Partial As files | Partial As files | Supported Templates, dry-run + apply | Supported Git-backed source dir | Supported Git history, diff + apply restore |
| agentsync | Supported MCP across agent targets | Supported Rules, skills, hooks + commands | Supported Dry-run + native/lossy/skipped report | Supported Canonical .agentsync tree | Supported Git checkpoints + revert |
| Ruler | Supported Central TOML servers, merge or overwrite | Partial Rules + skills, subagents experimental | Supported Native paths per agent, dry-run, skip warnings | Supported .ruler directory + ruler.toml | Partial Backup revert, subagents not revertible |
Evidence and operations
| Project | Receipts | Code impact | Agent activity | Cloud state | Learning / replay |
|---|---|---|---|---|---|
| BrigadeStack | Supported Command + Git + optional impact + run lineage | Supported Code impact, code-graph views, blast radius | Supported Center activity, fleet board, Command Deck, best-effort external harness sessions | Supported Cloud launch for Codex Cloud and Cursor, best-effort status, approved Grok Bot scout feed 0.32.0 beta Removes the Grok Bot queue and scout feed. Cloud launch for Codex Cloud and Cursor stays | Partial Manifest-scored promote + rollback, runs child/resume, outcome fork/diff |
| ActiveGraph | Partial Event trace + export | Not documented | Partial Behavior + tool events | Not documented | Supported Replay, fork + diff |
| Graphiti | Partial Episode provenance + validity windows | Not documented | Partial Optional OpenTelemetry operation traces | Not documented | Partial Temporal history + fact invalidation |
| Hindsight | Partial Recall traces + memory history | Not documented | Partial Memory lifecycle hooks + session tags | Not documented | Supported Reflective memory learning |
Scope
Every neighbor cell was checked on 2026-11-15 against the current official README, docs, or package page linked under Sources. Optional features may need configuration, and a packaged release can lag or differ from current docs. BrigadeStack cells describe version 0.31.0, the published release, checked on 2026-11-15 against the v0.31.0 tag linked below. A 0.32.0 beta line under a BrigadeStack cell says what the beta build changes. The status shown is still the 0.31.0 one.
Cloud state means tracking work that a cloud coding agent runs: launch, provider IDs, and status. Cloud-hosted memory or cloud embeddings do not count.
The evidence table’s Learning / replay column counts either reflective memory learning or execution replay. The memory table’s Learning column describes memory learning.
Care and operations counts documented memory edit, delete, history, and access controls. It does not require Brigade-style owner policy.
Provenance built only from IDs, timestamps, and change history is Partial. Explicit citations back to source observations are Supported.
Receipts, event traces, memory history, and data lineage answer different questions. A Partial receipts cell means a different kind of record, not a weaker command receipt.
BrigadeStack is a local file-first coordination layer for coding agents. The fleet hub, Center, Run View, MCP servers, and work dispatch --watch run in the foreground or under a service unit you install. brigadestack care install writes target-scoped systemd user timers or launchd agents, and on Windows it prints a Task Scheduler plan without creating tasks.
The work ledger is per-repo, machine-local JSON and is not a distributed task database. An optional fleet hub aggregates run events and arbitrates repo claims, and claims carry a generation so a stale operation is refused. The hub can also hold an opt-in Worklore work database. Its task dependencies are stored only and not yet enforced, and nothing consumes the queue. The fleet board can show external harness sessions. Local journals stay authoritative. An approved Grok Bot scout feed enqueues labeled GitHub issues only after --apply.
brigadestack work dispatch launches a wave of ready tasks from one repo into separate worktrees, with a worker cap, a session deadline, a launch ceiling, and --watch to refill capacity as workers finish. It does not dispatch across repos: campaign waves are for planning.
Ordinary runs have no hard run-budget. A run started with a declared budget gets hard wall-clock and worker-dispatch ceilings, and BrigadeStack refuses new dispatch past them. Cancelling workers already in flight is best-effort. Model, tool, token, and cost dimensions stay observed unless an adapter owns an enforcement boundary.
Campaigns aggregate ready work across repos and compose parallel-safe waves at query time from each member's per-repo partition. Center web views, the hub Command Deck, and the fleet board are read-only for ordinary viewers. Center CLI commands can write local reports, action records, and work imports. The hub roster page lets an admin save seat and routing settings. An administrator can also grant a named login a scoped operator session that sends an alert repair prompt to a registered destination.
brigadestack run cloud launch can start Codex Cloud or Cursor Cloud work from a bounded prompt file. The provider runs that work. BrigadeStack records the provider IDs and reports status as best-effort observations. External activity observations are best-effort too.
Memory care scans cards and queues review work. In 0.31.0 it does not rewrite card text: only memory care backfill --apply adds derived frontmatter, with a receipt. Safe targeted handoffs file automatically under owner policy, and ambiguous or risky notes wait for review. Reflection proposals built from successful manifest-backed outcomes wait for review by default. A hub policy, off by default, lets an allowlisted reflect principal apply two low-risk operations on its own: appending evidence and merging exact duplicates.
Outcome learning scores only receipts from verify manifests tracked under verify/manifests/. Ad-hoc --command and --argv-json receipts are audit-only evidence for a human reader. Learning is scored verification plus reviewed reflection proposals, not autonomous reflective memory learning.
BrigadeStack is not a hosted memory service, automatic release bot, or full autonomous fleet runtime. It does not publish packages. It pushes to GitHub only through run cloud land, which opens a pull request for review, and it sends chat messages only through a send action.
BrigadeStack sources, pinned to the v0.31.0 source commit: Comparison notes and scope, MCP sync, Execution model, Outcome scoring, Run budgets and cloud launch, Operator Center, Fleet hub, Scheduled care, Memory care, Hub memory and reflection, Beads import.
What the 0.32.0 beta changes
The 0.32.0 beta (build 0.32.0.dev20261114 on PyPI) removes the Grok Bot paths: the private job queue, the scout and build feeds that turned labeled GitHub issues into jobs, the workers, and the fleet-steward, Obsidian-operator, and T3-dispatch packs. The old commands print a removal message for one release. Client-side message encryption and the fleet keys commands are gone too. Plain hub messaging is unchanged. In their place, brigadestack hub connector setup --id cloud-agent gives any cloud agent that speaks MCP shared memory search, card reads, and handoff submission.
Verification gains observe and off modes beside the default strict mode. In observe mode, Claude Code and Codex hooks record the checks an agent runs as separate unsigned records, and brigadestack work checks reports first-pass rates by harness and model. These records are not receipts and do not score skills.
Memory care tends cards on its own. It applies derived date fixes, drains the review inbox, and moves notes that would overflow an instruction file such as USER.md into private cards. Every change is journaled and can be reversed with memory care undo. It never deletes a card.
Unchanged in the beta: brigadestack run, work dispatch, campaigns, Worklore, Center, the Command Deck and fleet boards, run cloud launch, MCP sync, and skills.
Beta sources, on the main branch: Memory care tending, Cloud agent connector, Work modes and observed checks.
Sources
Official sources behind each neighbor row, checked 2026-11-15.
- Beads: README.
- Gas Town: README, Per-bead dispatch lock, Per-assignee lock. Locks are per bead and per assignee on one machine, not a distributed lease.
- nWave: README.
- ActiveGraph: README.
- Claude Code agent teams: Agent teams. Experimental and off by default. /resume and /rewind do not restore in-process teammates, and a session has one team.
- Mem0: How it works, MCP and memory operations, Memory history, Memory export, Dream, OpenMemory status comment. Dream is a Mem0 Platform feature, and pattern synthesis needs the Pro plan. A 2026-07-17 comment on the Mem0 repo says OpenMemory is no longer maintained. That applies to OpenMemory only. This row covers Mem0.
- Letta Code: Letta Code README, Memory filesystem, Memory configuration. letta-ai/letta now points to letta-code. Memory blocks and archival passages remain documented only in the legacy V1 SDK docs, so this row covers the current memory filesystem.
- agentmemory: README, Changelog. Capture and learning depend on hooks, a running server, and a configured provider.
- Hindsight: CLI quickstart, Documents API, Recall API, Coding-agent hooks.
- Built-in harness memory: Claude Code memory, Codex memories, Cursor rules. A sample of three tools, not every coding agent. Cursor is covered by rules only: its docs list no automatic memory feature.
- OpenClaw: Memory architecture, Memory search, Memory CLI.
- add-mcp: README.
- Rulesync: Supported tools and features, CLI commands.
- chezmoi: Command overview, Templating, Quick start, Global flags (--dry-run).
- Ruler: README. Ruler labels itself a beta research preview. Subagent propagation has no documented revert.
- agentsync: Capability matrix, Rollback, CLI reference, Repository. The project labels itself beta.
- Graphiti: README, OpenTelemetry tracing. Graphiti only. The hosted Zep product is not counted.