Memory

Memory is what separates DAEMI agents from stateless chat interfaces. Every agent has two persistent memory layers, both managed by DAEMI's memory store.

---

Two Memory Layers

Short-Term Memory (STM)

STM is the agent's active working memory for the current session. Every message, tool result, and reasoning step is written here as the conversation unfolds. STM is:

• Semantically indexed — DAEMI's semantic indexing lets the agent retrieve relevant context from earlier in a long session without blowing the context window
• Session-scoped — starts fresh each session (though LTM primes it with relevant past context at session start)
• Bounded — the stm_window setting caps how many messages stay in the active context; older messages are still stored and can be retrieved by semantic search

Long-Term Memory (LTM)

LTM is persistent memory that survives session boundaries. It accumulates over time and is the basis of the agent's "knowledge of you." LTM is:

• Distilled from STM — at session end, DAEMI processes the session and writes significant items to LTM according to the agent's transfer_policy
• Queried at session start — before the conversation begins, DAEMI searches LTM for context relevant to the user's first message and injects it
• Grows indefinitely — there's no automatic pruning (you can manually clear it via the API)

---

How Memory Flows

Session Start
     │
     ▼
[LTM Query] ──── relevant past context ────► system prompt injection
     │
     ▼
[Conversation] ──── each turn ────► STM write
     │
     ▼
Session End
     │
     ▼
[STM → LTM distillation] per transfer_policy

---

Transfer Policies

Policy	What gets written to LTM
carry_all	All significant exchanges, decisions, and facts
facts_only	Only explicit facts and stated preferences
explicit_only	Only items the agent was directly told to remember
none	Nothing — the session is fully ephemeral

Set the policy in the agent's profile:

memory:
  transfer_policy: carry_all

---

Tuning Memory Behavior

memory:
  transfer_policy: facts_only
  ltm_max_results: 10    # max LTM records injected at session start
  stm_window: 50         # max messages in active STM context

ltm_max_results — higher values give the agent more past context but increase prompt size. Default: 10.

stm_window — how many recent messages stay in the live context window. Older messages are still stored and can be retrieved via semantic search when relevant. Default: 50.

---

Querying Memory via API

You can write to an agent's long-term memory directly:

# Write a fact to LTM
POST /api/agents/{agent_id}/memory/ltm
{
  "content": "User prefers async/await over Promise chains in Node.js",
  "type": "fact"
}

Response shape:

{
  "id": "mem_abc123",
  "content": "User prefers async/await over Promise chains in Node.js",
  "type": "fact",
  "created_at": "2026-02-14T10:30:00Z"
}

---

Clearing Memory

# Clear all LTM (irreversible)
DELETE /api/agents/{agent_id}/memory/ltm

caution

Clearing LTM is permanent. There's no undo. The agent loses everything it has learned across all past sessions.

---

Go Deeper

→ Profiles

Set transfer_policy and memory tuning options.

→ API Reference

All memory endpoints documented.