Memory (Hippocampus)

A 5-layer persistent memory system that gives asyncat continuity across sessions.

Memory layers

The Hippocampus models biological memory architecture with five distinct layers, each serving a different purpose in maintaining the agent's knowledge and context.

Layer	What it stores	Lifetime
Episodic	Session logs and conversation history — what happened, when, and in what order.	Persists across sessions, pruned after configurable retention period.
Semantic	Facts, learned knowledge, and persistent information — user preferences, project context, learned terminology.	Permanent until explicitly deleted.
Procedural	Learned skills and "how to" knowledge — step sequences for repeated tasks the agent has already mastered.	Permanent; reinforced through repeated use.
Working	Current context and active task state — goals, recent findings, pending decisions. The in-session scratchpad.	Cleared at session end.
Emotional	Tone preferences, communication style signals, and relationship history. Tracks whether the user prefers brief answers, formal language, etc.	Persists; adjusts incrementally over time.

How it works

On every message, the agent simultaneously reads from and writes to the relevant memory layers. On startup, memories are retrieved and injected into the context window, giving the agent a coherent sense of history.

# on session start — memory is loaded:
  episodic  → recent session summaries
  semantic  → known facts and user preferences
  procedural → frequently used skill sequences

# during a run — memory is updated:
  episodic  → log this interaction
  semantic  → extract and store new facts
  emotional → adjust tone model

Memory retrieval is semantic — the agent searches by meaning, not keyword. Relevant memories surface automatically based on the current task context.

Memory tools

The agent uses these tools to read and write memory explicitly during a run. See Workspace tools for the full reference.

Tool	Description
`save_memory`	Save a fact or piece of information to long-term memory with optional tags.
`recall_memory`	Retrieve memories relevant to a query using semantic search.
`list_memories`	List all stored memories, optionally filtered by type or tags.
`forget_memory`	Permanently delete a memory entry by ID.