Memory

JamJet Cloud bundles a hosted memory store so your agents can write facts and recall them later — across an entire fleet, scoped to a single agent, or scoped to an end-user. Embeddings are pgvector-backed; you bring your own provider key (BYOK), the cloud encrypts it at rest.

If you already use the open-source jamjet-engram library locally, the cloud surface is API-compatible: swap base_url to point at JamJet Cloud and the same add / recall / forget calls work against managed storage instead of your local SQLite.

When to use this

• Across an agent fleet. A support agent and a billing agent both need to know "User Y prefers formal greetings" — write the fact once, both recall it.
• Per end-user context. Tag facts with end_user_id so personalization survives across sessions without you running your own vector store.
• Right-to-erasure. A single forget(end_user_id="...") call soft-deletes every fact about that user, with one audit-log entry that records the count.

If you're building a single-agent app with no fleet and no per-user state, you probably don't need cloud memory yet. Local OSS Engram is enough; come back when you outgrow it.

How it's scoped

Every fact has three optional dimensions: project, agent, end-user. A recall with explicit filters returns the inclusive cascade — facts compatible with the filters, not strictly equal to them.

Worked example: agent X recalls about user Y. The query returns:

Omitted scopes mean "applies broadly." Explicit scopes narrow.

Setup

1. Configure the embedding provider

Cloud memory uses your own OpenAI key for embeddings — JamJet does not bill you for embedding cost. In the dashboard, go to Settings → Memory (or visit /dashboard/settings/memory directly) and:

1. Paste your OpenAI API key (sk-…).
2. Click Test connection to verify it works against text-embedding-3-small.
3. Click Save. The key is encrypted with AES-256-GCM and stored in projects.embedding_provider_key_encrypted as ciphertext bytes. Plaintext never crosses the DB boundary.
4. Toggle Memory enabled on (Save does this automatically on first set).

2. Test the endpoints

Three core operations: add, recall, forget. Plus read-side helpers facts and stats used by the dashboard.

# Add a fact
curl -X POST https://api.jamjet.dev/v1/memory/add \
  -H "Authorization: Bearer $JAMJET_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "content": "User likes formal greetings",
    "agent_id": null,
    "end_user_id": "u_alpha"
  }'
# → { "id": "abc-...", "content": "...", "created_at": "..." }

# Recall
curl -X POST https://api.jamjet.dev/v1/memory/recall \
  -H "Authorization: Bearer $JAMJET_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "how should I greet the user?",
    "end_user_id": "u_alpha",
    "k": 5
  }'
# → { "facts": [{ "id": "...", "content": "...", "similarity": 0.87, ... }] }

# Forget by id
curl -X POST https://api.jamjet.dev/v1/memory/forget \
  -H "Authorization: Bearer $JAMJET_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "id": "abc-...", "reason": "user request" }'
# → { "count": 1 }

import os
import httpx

API = "https://api.jamjet.dev"
HDR = {"Authorization": f"Bearer {os.environ['JAMJET_API_KEY']}"}

# Add a fact
r = httpx.post(f"{API}/v1/memory/add", headers=HDR, json={
    "content": "User likes formal greetings",
    "end_user_id": "u_alpha",
})
fact_id = r.json()["id"]

# Recall
facts = httpx.post(f"{API}/v1/memory/recall", headers=HDR, json={
    "query": "how should I greet the user?",
    "end_user_id": "u_alpha",
    "k": 5,
}).json()["facts"]

# Forget by id
httpx.post(f"{API}/v1/memory/forget", headers=HDR, json={
    "id": fact_id,
    "reason": "user request",
})

from jamjet.engram import EngramClient

client = EngramClient(
    base_url="https://api.jamjet.dev",
    api_key=os.environ["JAMJET_API_KEY"],
)

await client.add(content="User likes formal greetings", end_user_id="u_alpha")
facts = await client.recall(query="how should I greet?", end_user_id="u_alpha", k=5)
await client.forget(facts[0]["id"], reason="user request")

GDPR — delete by end-user

Right-to-erasure works in one call:

curl -X POST https://api.jamjet.dev/v1/memory/forget \
  -H "Authorization: Bearer $JAMJET_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "end_user_id": "u_alpha", "reason": "GDPR delete request 2026-05-12" }'
# → { "count": 47 }

Every fact tagged with that end_user_id (across all agents in the project) is soft-deleted in a single transaction. The audit_log gets one row recording the bulk delete with the count — not one row per fact, so erasing a chatty user doesn't flood your audit trail.

The deleted facts retain their content in the database with superseded_at set, but never appear in recall or facts again. Hard-delete-after-30-days is on the roadmap; for true compliance erasure today, contact support.

Tier limits

Memory storage is metered separately from traces. Counts are per-month:

Hitting the limit returns 402 Payment Required with dimension: "memory" in the response body. Upgrade in Settings → Billing.

The Memory dashboard tab shows your current count and tier-relative percentage at the top.

Browse facts in the dashboard

/dashboard/memory shows recent facts in the active project, with:

• Search box — empty shows newest-first; non-empty does a recall ranked by cosine similarity.
• Per-row forget — confirm modal, optimistic update.
• Stats header — count / quota / tier with a link to Settings → Memory for provider config.

The page is hidden when the project hasn't configured an embedding provider yet, with a CTA pointing to Settings.

Auth model

• API key auth (jj_…) — single project, scoped automatically. Use this from your agents.
• Dashboard auth (Supabase JWT) — scoped to projects you own. Used by the Memory tab and Settings page in the browser.
• Disabled project — when memory_enabled = false, all three endpoints return 403 with error_code: "memory_disabled". Toggling re-enables without touching data.

Endpoint reference

Full HTTP shapes are in the REST API reference. Quick summary:

Differences from open-source Engram

If you're using consolidate / context / search from OSS Engram today, those endpoints aren't in the cloud MVP. Open an issue with your use-case and we'll prioritize.