HeartBeatAgents

Agents That Remember Everything

View as Markdown

Three memory types work together to store conversations, accumulate knowledge, and retain learned behaviors across sessions, users, and even agent boundaries.

Three memory types

Episodic memory (conversation history)

Stores every interaction, with scoring that prioritizes important moments. Agents recall past interactions, reference previous decisions, and maintain continuity across sessions.

Semantic memory (facts & knowledge)

Builds a structured knowledge graph. When a user shares a preference or technical detail, the agent stores it as semantic memory and recalls it in future conversations.

Procedural memory (learned patterns)

Encodes successful strategies during task completion, so an agent applies learned behavioral patterns automatically the next time a similar task comes up.

Technical capabilities

  • Vector embeddings: semantic search backed by whichever embedding provider you connect, so recall matches by meaning rather than exact words. Without one connected, memory falls back to text-only matching automatically instead of failing. See Embedding Provider for the supported providers and how to turn this on.
  • Importance scoring: a continuous 0.0–1.0 scale based on emotional significance, decision impact, and recurrence frequency.
  • Automatic decay: a configurable decay rate (default 0.995/day) prevents memory bloat and keeps the most relevant context surfacing first.
  • Cross-agent sharing: memory can be shared at the team or organization level, not just scoped to a single agent.

Why memory doesn't grow forever

Once an agent accumulates enough memories, similar ones are automatically clustered and merged into a single, consolidated memory rather than kept as separate near-duplicates. This runs on a schedule in the background, and each agent's memory is capped, with the least important entries pruned first once that cap is reached. The store stays useful as it grows instead of turning into an unfiltered log.