Memory System

Memory system (Graphiti)

🌍 English translation is auto-generated on release. See the French version for the current complete page.

WorkPilot AI uses Graphiti, a semantic graph memory system, to retain knowledge across sessions. Agents no longer need to re-discover your project on each task.

Memory System

WorkPilot AI's memory layer is built on Graphiti, a graph-based semantic memory store that lets agents retain insights across sessions rather than starting from scratch on every task. It's implemented in integrations/graphiti/ and configured through the Electron app's onboarding/settings UI — CLI users can enable it directly by setting GRAPHITI_ENABLED=true in .env-files/.env.

Beyond the knowledge graph itself, a separate Memory Search (mem-search) layer provides progressive, token-efficient retrieval over existing memories — specifically task_logger traces and patterns extracted by the Learning Loop. It answers "have we hit this before?" without forcing an agent to load every candidate record just to discard most of them:

Layer	Cost	Purpose
Index	~100 tokens, always built	Lightweight lookup across candidate memories
Timeline	A couple of lines per entry	Quick scan of the top matches
Detail	Full record, fetched by id	Complete context, only on demand

The index is held to a fixed token budget by dropping entries (never by truncating what it keeps), and building it never reads a full record body — so the retrieval cost stays proportional to what's actually used. This progressive-disclosure approach is exposed to agents via skills/tooling/mem-search/.

Complementing both is the Memory Manager agent, which handles Graphiti knowledge-base upkeep, and the broader Learning Loop, which extracts patterns from completed builds and A/B-replays promoted changes — so improvements are backed by observed runs rather than guesses.

Together, these systems mean an agent working on a task can draw on relevant history — prior bugs, patterns, and decisions — without re-deriving context every session, while keeping retrieval costs proportional to what's genuinely needed.

---

➡️ Continue: Architecture