FocalPoint — AI Cognitive Operating System

Your AI forgets everything between conversations. FocalPoint fixes that.

Not just memory — cognitive infrastructure. FocalPoint tracks your projects, prepares context before tasks, uses a Three-Province review system for quality decisions, and proactively alerts you about stuck work.

The Problem

AI agents are stateless. Every conversation starts from zero — no memory of your projects, decisions, or progress. You waste time re-explaining context and manually tracking what's stuck.

Existing solutions only go halfway:

Tool	What it does	What it doesn't do
Mem0 / Zep	Remembers conversations	Doesn't track tasks or alert you
LangGraph / CrewAI

Orchestrates agents | No persistent cognitive layer |
| Claude / OpenAI memory | Remembers preferences | Doesn't manage work or deadlines |

They remember what was said. FocalPoint manages what needs to be done.

What You Get

• - Structured memory — Goal > Project > Milestone > Task hierarchy with status lifecycle
• Workbench — One call prepares goal, knowledge, context, subtasks, and role prompt
• Proactive alerts — Heartbeat detects blocked, stale, and at-risk tasks automatically
• Knowledge documents — Attach design docs to nodes; child tasks inherit parent knowledge
• Role-based thinking — Strategy, Review, and Execution roles see filtered perspectives
• Three-Province review — Parallel review by two reviewers before execution; max 3 rejections then escalate
• Full-text search — FTS5 search across titles, narratives, and knowledge documents
• GitHub + Notion sync — Issues and pages auto-sync as FocalPoint nodes
• Zero dependencies — Runs 100% locally on SQLite. No vector DB, no Redis, no cloud.

Competitive Comparison

Capability	Mem0	Zep	Letta	CrewAI	Claude	FocalPoint
Persistent memory	Yes	Yes	Yes	Yes	Yes	Yes
Task lifecycle management

- | - | - | Partial | - | Yes | | Dependency graph (DAG) | - | - | - | - | - | Yes | | Proactive alerts | - | - | - | - | - | Yes | | Knowledge docs + inheritance | - | - | - | - | Partial | Yes | | Role-based context | - | - | - | Partial | - | Yes | | Decision review workflow | - | - | - | - | - | Yes | | Full-text search | Vector | Vector | Vector | - | - | FTS5 | | MCP native | - | - | - | - | Proprietary | Yes | | Self-hosted, zero deps | Partial | Partial | Yes | Yes | - | Yes |

No competitor combines structured task management + proactive alerts + knowledge inheritance + role-based context + review workflow in a single MCP-native package.

Setup

1. Install

CODEBLOCK0

2. Add MCP server

CODEBLOCK1

3. Restart OpenClaw

That's it. 22 tools are now available.

How It Works

Architecture: Brain-Spine Model

CODEBLOCK2

• - Brain = LLM. Only reads context and issues Tool Calls.
• Spine = Deterministic engine. All logic here. LLM never touches storage directly.

Storage: Pure SQLite

CODEBLOCK3

No vector database. No Redis. No PostgreSQL. One SQLite file is everything.

Work Mode

Workbench — prepare before you work

CODEBLOCK4

Three Roles

Role	Focus	Sees
Strategy (Maker)	Should we do this? Priority?	Decisions + feedback
Review (Reviewer)

Any risks? Historical lessons? | Risk notes + progress | | Execution (Engineer) | How to build it? Acceptance criteria? | Technical details + progress |

Same data, different thinking modes. The role prompt guides the AI's perspective.

Three-Province Review — quality decisions

For major decisions (new features, architecture changes, tech choices):

CODEBLOCK5

Knowledge Documents — persistent design context

CODEBLOCK6

Log Categories

CODEBLOCK7

Different roles see different categories. Strategy sees decisions + feedback. Execution sees technical + progress.

Use Cases

Project tracking

CODEBLOCK8

Decision memory

CODEBLOCK9

Risk detection

CODEBLOCK10

Full-text search

CODEBLOCK11

Available Tools (22)

Write (11)

Tool	What it does
INLINECODE0	Create a project/task/goal/milestone
INLINECODE1

Change status (inbox/active/waiting/done/dropped) |

| update_field | Update title, summary, deadline, etc. | | attach_node / detach_node | Move tasks in hierarchy | | add_dependency / remove_dependency | Manage task dependencies | | append_log | Record decisions, progress, risks (with category) | | unarchive | Restore completed/dropped tasks | | set_persistent | Protect tasks from auto-archive | | set_knowledge | Attach knowledge documents to nodes |

Read (5)

Tool	What it does
INLINECODE11	Get full details of a work item
INLINECODE12

Find tasks by filters or full-text search |

| get_knowledge | Read knowledge with parent inheritance | | delete_knowledge | Delete a knowledge document | | get_assembly_trace | Debug context assembly |

Cognitive (4)

Tool	What it does
INLINECODE16	Load memory context (call at conversation start)
INLINECODE17

Scan for risks: blocked, stale, at-risk tasks |

| activate_workbench | Prepare working context with role + knowledge | | get_context_bundle | Get role-filtered, token-budgeted context |

Review (1)

Tool	What it does
INLINECODE20	Three-Province parallel review

Runtime (1)

Tool	What it does
INLINECODE21	Switch AI attention to a specific task

Automatic Memory Rules

Follow these rules in EVERY conversation:

1. 1. Conversation start -> Call bootstrap to load memory
2. Before starting a task -> activate_workbench to prepare context
3. User makes a decision -> append_log with category="decision"
4. Risk identified -> append_log with category="risk"
5. Task progresses -> INLINECODE26
6. Design conclusions -> set_knowledge to persist for future sessions
7. Before conversation ends -> append_log key takeaways
8. Every ~10 min -> heartbeat to check for risks

Stats

Metric	Value
Tests	665
MCP Tools

22 | | External dependencies | 0 (pure SQLite) | | Cold start | < 100ms | | Supported LLMs | Any (via MCP protocol) | | Platforms | Claude Desktop, OpenClaw, any MCP client |

Requirements

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

Links

• - GitHub
• PyPI