Agent Memories

A memory is one short fact an agent was told to keep, or a rule it was told to follow. These endpoints are the operator's view of that store: see exactly what is remembered, teach a fact or a rule directly, correct one, turn one into the other, and delete anything that should not be there.

The injection window: what the agent reads without asking

Every memory stays searchable forever. The newest of each kind are handed to the agent at the start of a turn, so it acts on them without deciding to look anything up: up to 100 rules, 500 facts you or the agent saved, and 150 facts picked up from chats, each kind in its own window. That set is the injection window, and injected: true marks it on every listed memory.

The window is applied today on the linked personal WhatsApp channel. On every other channel a memory is still remembered and still found by recall_memory; it is simply not placed in the prompt ahead of time, so injected there tells you where a fact sits in the order rather than that it was already read.

So a fact does not stop existing when it falls out of the window. It stops being free. If a preference must hold on every single reply, keep it near the top by restating it: memories are ordered newest first, and a fresh statement of the same preference moves it back into the window.

Facts and rules

Every memory has a memory_type.

Write a rule as one instruction, the way you would say it out loud. A fact can be made a rule and a rule moved back to a fact with PATCH (below), without deleting and re-teaching it.

What is listed here

A memory is stored in one of two forms, and this surface serves only the first.

Transcripts are deliberately absent: a transcript is not something an operator can meaningfully review or correct one line at a time, so it stays retrieval-only rather than becoming a browsable message archive.

Credential class: dashboard JWT or a cm_ key

Authorization: Bearer cm_your_api_key

A memory is an attribute of a bot rather than a resource class of its own, so it reuses the bots:* scopes and adds none.

Base URL for every example: https://api.callmissed.com. Errors are {"detail": "..."}.

The memory object

{
  "id": "7c9e6679-7425-40de-944b-e07fc1f90ae7",
  "fact": "Wholesale orders ship Tuesdays only. Never promise a Friday dispatch.",
  "created_at": "2026-09-08T12:00:00Z",
  "injected": true,
  "captured_by": "remember_tool",
  "memory_type": "fact"
}

Memories belong to one agent

A memory is scoped to a workspace and a bot. Two agents in the same workspace do not share a fact, and re-pointing a phone number at a different agent gives it a different set of memories, not the same set under a new name. Move a fact by POSTing it to the second agent.

GET /api/v1/agent-memories

Everything one agent has been told to remember, newest first. Requires bots:read.

curl "https://api.callmissed.com/api/v1/agent-memories?bot_id=b1f2c3d4-5678-90ab-cdef-1234567890ab&limit=50" \
  -H "Authorization: Bearer cm_your_api_key"

{
  "bot_id": "b1f2c3d4-5678-90ab-cdef-1234567890ab",
  "memories": [
    {
      "id": "7c9e6679-7425-40de-944b-e07fc1f90ae7",
      "fact": "Wholesale orders ship Tuesdays only. Never promise a Friday dispatch.",
      "created_at": "2026-09-08T12:00:00Z",
      "injected": true,
      "captured_by": "remember_tool",
      "memory_type": "fact"
    }
  ],
  "max_injected": 500
}

max_injected is the size of the injection window. Read it instead of hardcoding 500, so a change to the window does not silently make your own display wrong.

POST /api/v1/agent-memories

Teach the agent a fact or a rule directly, without waiting for it to come up in a conversation. Returns 201 and the memory object. Requires bots:write, or an owner or admin JWT.

curl -X POST https://api.callmissed.com/api/v1/agent-memories \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "bot_id": "b1f2c3d4-5678-90ab-cdef-1234567890ab",
    "fact": "Wholesale orders ship Tuesdays only. Never promise a Friday dispatch."
  }'

The new memory always comes back with injected: true, because it is the newest, and captured_by: "console", which is how a taught fact is distinguished from one the agent chose to keep.

To teach a rule, send the same request with "memory_type": "rule":

curl -X POST https://api.callmissed.com/api/v1/agent-memories \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "bot_id": "b1f2c3d4-5678-90ab-cdef-1234567890ab",
    "fact": "Never share my home address.",
    "memory_type": "rule"
  }'

PATCH /api/v1/agent-memories/{memory_id}

Correct a memory's text, make a fact a rule or move a rule back to a fact, or both at once. Returns 200 and the updated memory object. Requires bots:write, or an owner or admin JWT.

Send at least one of the two.

curl -X PATCH https://api.callmissed.com/api/v1/agent-memories/7c9e6679-7425-40de-944b-e07fc1f90ae7 \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{"memory_type": "rule"}'

To correct the wording instead:

curl -X PATCH https://api.callmissed.com/api/v1/agent-memories/7c9e6679-7425-40de-944b-e07fc1f90ae7 \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{"fact": "Wholesale orders ship Tuesdays and Thursdays."}'

Either change keeps the memory's created_at, so it stays where it was in the order: correcting a typo never moves a memory into or out of its window. Corrected text is re-indexed in the same step; if that fails, nothing changes.

DELETE /api/v1/agent-memories/{memory_id}

Forget one fact or rule. Returns 204 and no body. Requires bots:write, or an owner or admin JWT.

curl -X DELETE https://api.callmissed.com/api/v1/agent-memories/7c9e6679-7425-40de-944b-e07fc1f90ae7 \
  -H "Authorization: Bearer cm_your_api_key"

Deletion is idempotent: 204 whether the memory was there or not, so a retry after a dropped connection is safe. It is also scoped to memories, so this route cannot be used to remove an ordinary knowledge document by guessing at its id. The fact and its search index go together; there is no way to read a deleted memory back.

Errors

Billing

A memory is embedded when it is written, so that the agent can find it later by meaning rather than by keyword. That embedding is billed like any other embedding on your account, at the rate on the pricing page. Corrected text is re-embedded and billed the same way. Listing and deleting are free.

A memory whose embedding did not complete still appears in the list, with fact falling back to its first 120 characters, because an un-embedded memory is exactly the thing you need to see rather than one that quietly vanishes.