focalpoint 0.3.3

FocalPoint — The attention manager for AI agents. Persistent memory, structured work tracking, proactive risk alerts, and bidirectional sync with GitHub/Notion.

FPMS — Focal Point Memory System

The attention manager for AI agents. FPMS gives your AI persistent memory, structured work tracking, and proactive risk alerts — so it never forgets what it's working on.

pip install fpms

The Problem

AI agents forget everything between conversations. Even with 1M token context windows, they can't:

• Remember what they were working on yesterday
• Notice that a task has been stuck for 3 days
• Prioritize which information to load when context is limited
• Track projects across multiple conversations

Existing memory tools (Mem0, Zep) store conversation snippets. But remembering what was said is not the same as managing what needs to be done.

What FPMS Does

FPMS is a cognitive engine that manages your AI's attention:

1. REMEMBER   — Persistent work items across conversations
2. FOCUS      — Token-budgeted context loading (only load what matters)
3. ALERT      — Proactive heartbeat: "Task X is blocked for 3 days"
4. STRUCTURE  — Projects → Milestones → Tasks with dependencies

How It Compares

Capability	Mem0	Zep	Letta	OpenViking	FPMS
Conversation memory	Yes	Yes	Yes	Yes	Via prompt rules
Token budget management	No	No	Yes	Yes	Yes (L0/L1/L2/L_Alert)
Structured work items	No	No	No	No	Yes
Parent-child hierarchy	No	No	No	No	Yes (unlimited depth)
Task dependencies	No	No	No	No	Yes
Proactive risk alerts	No	No	No	No	Yes (heartbeat)
Anti-Amnesia	No	No	No	No	Yes
Status lifecycle	No	No	No	No	Yes (inbox/active/waiting/done)
Auto-archive	No	No	No	No	Yes
External sync (GitHub + Notion)	No	No	No	No	Yes (bidirectional)
Write-back to external tools	No	No	No	No	Yes

Other tools remember what was said. FPMS manages what needs to be done.

Who Is FPMS For?

Solo founders & indie hackers

You juggle 10 projects across multiple AI conversations. FPMS keeps track so your AI always knows the full picture.

Example: You tell Claude "let's work on the auth system" — FPMS loads all related context: the task was created last week, you decided on JWT over sessions, there's a blocking dependency on the user model.

AI agent builders

You're building autonomous agents that run for hours. Without FPMS, they lose track of priorities and forget completed work.

Example: Your agent runs a heartbeat scan every 10 minutes and discovers: "Task 'Deploy to staging' has been active for 2 days with no progress. 'Fix login bug' is blocked by 'Update user model' which is still in inbox."

Teams using AI assistants

Multiple people interact with AI on the same project. FPMS provides a shared memory layer so the AI has continuity regardless of who's talking to it.

Example: Alice tells Claude to create a new API endpoint. Bob asks Claude about project status the next day — FPMS shows the endpoint task is active and 60% complete.

Use Cases

1. Project Management with AI Memory

You: "Create a project for the Q2 product launch"
AI:  → create_node(title="Q2 Product Launch", node_type="project", is_root=True)

You: "Add tasks: redesign landing page, setup analytics, write launch email"
AI:  → creates 3 child task nodes under the project

(Next day, new conversation)
You: "What's the status of the launch?"
AI:  → bootstrap() loads context
     → "Q2 Product Launch: 3 tasks, 0 completed. 'Redesign landing page'
        has been in inbox for 2 days — should I activate it?"

2. Decision Tracking

You: "We're going with Stripe instead of PayPal for payments"
AI:  → append_log(node_id="payments-task", content="Decision: Stripe over PayPal.
       Reason: better API, lower fees for international transactions")

(Two weeks later)
You: "Why did we pick Stripe again?"
AI:  → get_node() → reads decision log
     → "You chose Stripe on March 15 because of better API and lower
        international fees"

3. Risk Detection

(Agent runs heartbeat automatically)
AI:  → heartbeat() detects:
     - "Deploy to prod" BLOCKED for 4 days (dependency on code review)
     - "Update docs" STALE — no activity for 7 days
     - "Fix memory leak" AT RISK — deadline is tomorrow, still active

AI:  → "I found 3 issues: a blocked deploy, stale docs, and an at-risk
        memory leak fix due tomorrow. Want me to help prioritize?"

4. GitHub Integration

Your GitHub issues automatically sync into FPMS:

  GitHub Issue #42 "Add dark mode"  →  FPMS node (synced)
  GitHub Issue #43 "Fix crash"      →  FPMS node (synced)

AI can now reason across your codebase AND your task list:
  "Issue #42 has been open for 2 weeks and assigned to you.
   The related PR was merged yesterday — should I close it?"

5. Cross-Conversation Continuity

Conversation 1 (Monday):
  You: "Start working on the API refactor"
  AI:  → creates node, status: active

Conversation 2 (Wednesday):
  You: "Continue where we left off"
  AI:  → bootstrap() → "You started the API refactor on Monday.
         Last log: refactored /users endpoint, /orders is next."

Conversation 3 (Friday):
  You: "How's the refactor going?"
  AI:  → heartbeat() → "API refactor is active, 2/5 endpoints done.
         At current pace, you'll finish next Wednesday."

Quick Start

Option 1: MCP Server (recommended)

Works with Claude Desktop, Claude Code, and OpenClaw.

pip install fpms
fpms  # starts MCP server on stdio

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "fpms": {
      "command": "fpms"
    }
  }
}

OpenClaw (openclaw.yaml):

mcp_servers:
  fpms:
    command: fpms

Option 2: Python API

from fpms.spine import SpineEngine

engine = SpineEngine(db_path="./data/fpms.db")

# Cold start — load memory context
bundle = engine.bootstrap()

# Create and manage work items
engine.execute_tool("create_node", {
    "title": "Ship MVP", "node_type": "project", "is_root": True
})

# Proactive risk scan
alerts = engine.heartbeat()

Architecture

AI Agent (Claude / GPT / OpenClaw)
    |
    v  MCP Tool Call (stdio)
+----------------------------------+
|  mcp_server.py                   |  18 MCP tools (FastMCP)
|    +------------------------+    |
|    |    SpineEngine         |    |  Core engine
|    |  +---------+--------+  |    |
|    |  | Tools   | Heart- |  |    |
|    |  | Execute | beat   |  |    |
|    |  | Store   | Bundle |  |    |
|    |  | Valid   | Risk   |  |    |
|    |  +---------+--------+  |    |
|    |  Adapters (GitHub...)  |    |
|    +------------------------+    |
+----------------------------------+
    |
    v
  SQLite + Narrative Markdown

18 MCP Tools: create_node, update_status, update_field, attach_node, detach_node, add_dependency, remove_dependency, append_log, unarchive, set_persistent, shift_focus, expand_context, get_node, search_nodes, get_assembly_trace, heartbeat, bootstrap, get_context_bundle

Key Concepts

Concept	Description
Node	A work item (project, task, goal, milestone) with status lifecycle
Heartbeat	Periodic scan that detects blocked, stale, and at-risk nodes
Context Bundle	Token-budgeted payload (L0 dashboard + L1 neighborhood + L2 detail)
Anti-Amnesia	Mechanism that prevents AI from forgetting active work mid-conversation
Rollup	Child task completion automatically bubbles up to parent nodes
Narrative	Markdown log of decisions, events, and progress for each node

Requirements

• Python 3.10+
• No external services needed — runs 100% locally on SQLite

License

Business Source License 1.1 — Free to use, modify, and deploy. Cannot be used to build a competing commercial memory service. Converts to Apache 2.0 on 2030-03-22.

Links

• GitHub
• PyPI
• ClawHub
• Usage Guide