agendum
Project memory and scoping engine for AI coding agents.
AI coding agents are stateless — they forget between sessions, lose decisions, and have no way to scope complex work. agendum is an MCP server that gives any agent (Claude Code, Cursor, Windsurf, Cline, and others) persistent project state, bounded work packages, and cross-session continuity.
| Without agendum | With agendum |
|---|
| Agent forgets everything between sessions | Picks up exactly where it left off |
| No scope — agent modifies random files | Bounded work packages with file lists, acceptance criteria, constraints |
| Decisions lost — same mistakes repeated | Decisions and patterns persist in searchable memory |
| No task ordering — agent picks randomly | Dependency graph with auto-unblocking and priority scoring |
| Learning locked inside one project | Cross-project learnings carry patterns forward |
Quick Start
pip install agendum
claude mcp add agendum -- uvx agendum --home serve
Works with any MCP client — see setup for Cursor, Windsurf, VS Code, and others.
How It Works
flowchart LR
A["PLAN\nwrite plan file"] --> B["pm_ingest\nboard items + deps"]
B --> C["pm_next\nwork package + context"]
C --> D["EXECUTE\nagent implements"]
D --> E["pm_done\ndecisions + patterns"]
E -->|next task| C
E -->|new session| F["pm_status\nresume context"]
F --> C
Each pm_done records decisions and patterns that enrich future pm_next calls — context compounds across sessions.
Example session:
You: I have a plan file for the API rewrite. Ingest it.
Agent:
→ pm_ingest(project="api-rewrite", plan_file="plan.md")
Ingested 4 board items from plan:
item-001: Schema design [high]
item-002: Resolver layer (depends on item-001)
item-003: Auth middleware (depends on item-001)
item-004: Integration tests (depends on item-002, item-003)
You: What should I work on?
Agent:
→ pm_next(project="api-rewrite")
Work package for item-001 "Schema design":
Context: project rules, memory from last session
Scope: Define GraphQL schema types
Acceptance criteria: Types for User, Product, Order
You: Done with the schema. Here's what I decided...
Agent:
→ pm_done(project="api-rewrite", item_id="item-001",
decisions="Using code-first with Strawberry",
patterns="N+1 queries need DataLoader",
verified=True)
Marked item-001 as done. Unblocked: item-002, item-003
> Next: pm_next("api-rewrite") to continue with newly unblocked tasks
Setup & Orientation
| Tool | Purpose |
|---|
pm_init | Initialize board directory (optional — auto-initializes on first use) |
pm_project | Create, list, or get projects |
pm_status | Dashboard — item counts, recent progress, memory health, suggested next task |
Planning & Backlog
| Tool | Purpose |
|---|
pm_add | Add an item with type, priority, tags, dependencies, acceptance criteria |
pm_board | View and filter the project board |
pm_ingest | Import a Markdown plan file into bounded board items with dependencies |
Execution Loop
| Tool | Purpose |
|---|
pm_next | Get the next scoped work package with complexity signal and enriched context |
pm_done | Complete an item — record decisions, patterns, learnings; auto-extract from git; auto-unblock dependents |
pm_block | Report a task as blocked with reason |
Knowledge & Search
| Tool | Purpose |
|---|
pm_memory | Read, write, append, or search project memory (decisions, patterns, project knowledge) |
pm_learn | Record global or project-scoped learnings with tags and topic entities |
pm_search | Hybrid search across all knowledge — memory, learnings, completed items |
pm_consolidate | Clean memory corruption, deduplicate learnings, detect contradictions |
pm_supersede | Soft-invalidate a learning — excluded from all future searches |
Hybrid Search
pm_search combines three signals to find relevant knowledge across memory, learnings, and completed board items:
- FTS5 with Porter stemming —
auth matches authentication, config matches configuration. Always on, zero config.
- Vector search (optional) — Install
agendum[vectors] to add semantic similarity via fastembed + sqlite-vec. Activates automatically alongside FTS5.
- Entity graph — Topics and tags form a knowledge graph. Entries sharing 2+ entities are linked automatically. Graph expansion surfaces related knowledge that keyword search misses.
All three signals are fused via Reciprocal Rank Fusion (RRF), then reranked by recency and access frequency. The index rebuilds from Markdown files — no data loss if it gets corrupted.
Key Capabilities
- Adaptive context budget — enrichment scales with task complexity: 4K chars for trivial tasks, up to 10K for large ones
- Verification gate —
pm_done(verified=True) distinguishes tested from untested completions
- Git auto-extract —
pm_done reads git diff and git log automatically when no files are specified
- Pluggable enrichment pipeline — four context sources injected into every work package: project rules (CLAUDE.md/AGENTS.md), memory, dependency context, learnings
- Dependency resolution — topological ordering with cycle detection; dependents unblock automatically when upstream tasks complete
- Memory health —
pm_status warns about corrupted entries; pm_consolidate strips XML fragments, deduplicates, and flags contradictions
- Zero config — auto-initializes on first tool call, derives board name from git remote
- Git-native storage — all state is human-readable Markdown + YAML in
.agendum/, diffable and committable
Installation
All MCP clients except VS Code use the same config. Add to the appropriate file:
{
"mcpServers": {
"agendum": {
"command": "uvx",
"args": ["agendum", "--home", "serve"]
}
}
}
| Client | Config location |
|---|
| Claude Code | Run: claude mcp add agendum -- uvx agendum --home serve |
| Cursor | .cursor/mcp.json in project root |
| Windsurf | ~/.codeium/windsurf/mcp_config.json |
| Cline | Settings › MCP Servers › Edit |
| Roo Code | MCP settings file |
| Claude Desktop | claude_desktop_config.json |
VS Code (GitHub Copilot): Uses "servers" instead of "mcpServers". Add to .vscode/mcp.json:
{
"servers": {
"agendum": {
"command": "uvx",
"args": ["agendum", "--home", "serve"]
}
}
}
CLI (standalone)
pip install agendum
agendum project create my-app
agendum status
agendum next my-app
Storage Layout
All state lives in ~/.agendum/ (or .agendum/ in your project if you prefer local storage):
~/.agendum/
├── .cache/
│ └── search.db # FTS5 + vector search index (auto-rebuilt)
├── config.yaml
├── projects/
│ └── webapp/
│ ├── project.yaml # Project metadata
│ ├── board/
│ │ ├── item-001.md # Markdown + YAML frontmatter
│ │ └── item-002.md
│ └── learnings/ # Project-scoped learnings
│ └── learning-001.md
├── learnings/ # Cross-project learnings
│ └── learning-001.md
└── memory/
├── decisions.md # Key decisions + rationale
└── patterns.md # Discovered conventions
Architecture
src/agendum/
├── server.py # MCP server wiring (FastMCP)
├── tools.py # 14 MCP tools
├── models.py # Pydantic models (BoardItem, WorkPackage, SearchResult)
├── task_graph.py # Dependency resolution + topological levels
├── config.py # Shared configuration
├── env_context.py # Git diff/log auto-extraction
├── cli.py # CLI interface
├── enrichment/
│ ├── pipeline.py # ContextEnricher, budget allocation
│ └── sources.py # ProjectRules, Memory, Dependency, Learnings sources
└── store/
├── board_store.py # BoardItem CRUD
├── board_format.py # Markdown <-> BoardItem serialization
├── project_store.py # Project metadata
├── memory_store.py # Scoped memory storage
├── learnings_store.py # Global and project-scoped learnings
├── search_index.py # FTS5 + vector + entity graph + RRF
├── embedding.py # Lazy fastembed wrapper (optional)
└── locking.py # get_lock() + atomic_write()
Development
git clone https://github.com/sralli/agendum.git
cd agendum
uv sync
uv run pytest tests/ -v
uv run ruff check .
uv run ruff format --check .
License
MIT