Agents that remember everything
Every conversation is mined for facts, decisions, and context. Extracted memories are stored in a structured knowledge graph — so your agents never start from zero.
How Memory Extraction Works
From conversation to contextual recall
Conversation
User chats with an agent about architecture decisions, task assignments, or strategy.
Extraction
Claude-local CLI parses the conversation and extracts atomic facts as structured YAML.
Storage
Facts are organized into PARA folders: Projects, Areas, Resources, Archive.
Graph
Relationships between facts form a knowledge graph for contextual retrieval.
Recall
Future conversations automatically surface relevant past memories for the agent.
PARA Knowledge Organization
Inspired by Tiago Forte's PARA method, extracted memories are organized into four top-level folders. This makes retrieval fast and contextually relevant.
fact:
id: "mem-2026-001"
type: "architectural_decision"
confidence: 0.94
content: |
Switched from JWT to session-based auth
for company-scoped API keys.
context:
project: "Auth Rewrite"
agent: "Claude"
date: "2026-05-21"
relations:
- type: "supersedes"
target: "mem-2025-089"
- type: "related_to"
target: "mem-2026-003"Contextual Recall in Action
When you return to a conversation, relevant memories surface automatically.
Let's continue the auth work we discussed last week.
Right — we switched from JWT to session-based auth for company-scoped API keys. PR #234 is merged and all 47 tests pass. The next step is updating the CLI to use the new session flow. Should I create a task for that?