Docs

ebbmem

Hosted memory for agents. Use MCP, the HTTP API, or a client SDK. This site does not publish a download.

Overview

ebbmem stores text and recalls it. Four tiers are accepted on remember: working, episodic, semantic, and procedural.

Three ways in:

  • HTTP API. Remember is POST /api/v1/memory/remember. Recall is POST /api/v1/memory/recall or GET /api/v1/memory/recall.
  • MCP. Streamable HTTP at POST /mcp. GET /mcp is the legacy SSE route.
  • Client SDKs for Python, TypeScript, Java, Go, C#, and Rust. The packages are not published yet.

The full operation list is the API reference, with the same document in Redoc and Swagger UI.

Quickstart

A remember call needs a hosted origin, a credential, and a JSON body whose required field is content. This page does not publish the public hostname. Set EBBMEM_ORIGIN to the origin you were given, with no trailing path. Set EBBMEM_TOKEN to a Bearer token.

curl -sS -X POST "${EBBMEM_ORIGIN}/api/v1/memory/remember" \
  -H "Authorization: Bearer ${EBBMEM_TOKEN}" \
  -H "Content-Type: application/json" \
  -H "X-Ebb-Namespace: default" \
  -d '{"content":"User prefers metric units","tier":"episodic","importance":0.8}'

A successful remember responds with HTTP 201 and JSON fields id, memory_id, and status.

curl -sS -X POST "${EBBMEM_ORIGIN}/api/v1/memory/recall" \
  -H "Authorization: Bearer ${EBBMEM_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{"query":"metric units","limit":5,"profile":"Balanced"}'

Recall JSON includes query, count, results, profile, and namespace. X-API-Key can replace the Authorization header. The same calls from a client SDK are on the SDK overview.

Authentication

Memory routes accept Authorization: Bearer or the header X-API-Key, and the namespace header X-Ebb-Namespace. The clients read EBBMEM_ORIGIN, EBBMEM_TOKEN, EBBMEM_API_KEY, and EBBMEM_NAMESPACE. Details are in auth and config.

A query token is accepted only on GET /api/v1/events/stream, for clients that cannot set Authorization. This site does not issue credentials.

Namespaces

Remember and recall take a namespace in the JSON body, or in the header X-Ebb-Namespace. The body value wins when both are set.

When neither is set, a caller whose tenant id is not default or system uses {tenant}:default. Otherwise the namespace is default.

A caller is limited to namespaces for their tenant. A mismatch returns 403. Client SDKs send EBBMEM_NAMESPACE as X-Ebb-Namespace.

Remember and recall

POST /api/v1/memory/remember requires content. Optional fields: tier (working, episodic, semantic, procedural), tags, importance, valence, arousal, namespace, metadata.

POST /api/v1/memory/recall and GET /api/v1/memory/recall require query. Optional fields: limit, profile (Balanced, FactualPrecision, CreativeExploration, DebugDiagnostic), min_score, domain, namespace.

Batch ingest is POST /api/v1/memory/remember/batch and requires items. Other memory routes (reinforce, dream, reflect, wander, introspect, status, update, delete, graph) are in the API reference.

MCP

The hosted MCP endpoint is POST /mcp on the same origin as the HTTP API. It speaks streamable HTTP JSON-RPC. GET /mcp is the legacy SSE route. Send Authorization: Bearer or X-API-Key.

Claude Desktop, Cursor, and VS Code remote MCP entries need that full URL plus the auth header. This site does not publish the public hostname, so it does not include a client config file. The same origin is used by the client SDKs and the HTTP API.

Errors and rate limits

Memory routes document these statuses:

  • 201 remember succeeded. Recall uses 200.
  • 400 constraint violation or malformed input.
  • 401 missing or invalid Bearer token or X-API-Key.
  • 403 role or namespace denial.
  • 404 the target was not found.
  • 422 validation error.
  • 429 the token bucket is empty.

Error JSON uses code and message. A 429 response sets Retry-After, X-RateLimit-Limit, and X-RateLimit-Remaining.

When a deployment does not set EBBMEM_RATE_LIMIT_CAPACITY or EBBMEM_RATE_LIMIT_RPS, the bucket capacity is 1000 and the refill is 200 tokens per second. Remember and recall cost 1. Remember batch costs 2. Dream, reflect, wander, and vacuum cost 5. A hosted deployment can set different limits. Those overrides are not published here.

Liveness is GET /healthz.

Benchmarks

Two QA figures are published, plus keyless LongMemEval retrieval. Each one names its models and its protocol. LongMemEval keyless measures whether the labeled evidence is in the top k. It is not answer accuracy.

MindSpan QA
94.0%
376/400. Dev 189/200 (94.5%), heldout 187/200 (93.5%). gpt-4o-mini answers, gpt-4o-mini LLM judge, temperature 0, k=15. Extraction: gemini-3.1-flash-lite. Embeddings: gemini-embedding-001. Dev 189/200 (94.5%), heldout 187/200 (93.5%).
LoCoMo QA
80.8%
1,244/1,540. 95% CI 78.7–82.7%. gpt-4o-mini answers, gpt-4o-mini LLM judge, temperature 0. Extraction: gemini-3.1-flash-lite. Embeddings: gemini-embedding-001. LoCoMo standard protocol. Neutral default soul, not a dataset persona. Neutral default soul, not a dataset persona.
LongMemEval, keyless

Retrieval. Whether the labeled evidence is in the top k.

96.8%

session hit@10

87.2%

turn hit@10

Retrieval only, on a 100-question sample of LongMemEval-S (cleaned). 94 questions with labeled evidence scored. This is not answer accuracy.

Zero LLM calls. Zero API keys.

Embeddings: all-MiniLM-L6-v2 (ONNX, 384-d). Entity extraction: GLiNER small v2.1 (ONNX).

kEvidence session in top-kEvidence turn in top-kEvidence turn recall
594.7%80.9%66.7%
1096.8%87.2%79.9%
2097.9%94.7%90.0%
50100%96.8%93.7%
Fully local: no LLM anywhere. Disk-backed store, reopened before recall.

MindSpan QA by category

All 17 categories from the combined run, sorted by question count. Question answering only.

CategoryCorrect%
Family life51/6085.0%
Craft mastery41/41100.0%
Episodic precision36/3894.7%
Career decisions33/33100.0%
Social dynamics30/3196.8%
Early childhood28/3093.3%
Daily living29/29100.0%
Home maintenance24/2788.9%
Eldercare logistics21/2295.5%
Health wellness21/2295.5%
Logistical grounding16/1794.1%
Kinship heritage14/1687.5%
Enterprise architecture12/12100.0%
Temporal anchors12/12100.0%
Biographical chronology5/683.3%
Community engagement1/250.0%
Financial decisions2/2100.0%
Retrieval · hit@15
95.5%

382/400. At least one relevant memory in the top 15. Not answer accuracy.

Retrieval · recall@15
89.2%

463/519 relevant records. Not answer accuracy.

Retrieval, not answer accuracy. hit@15 means at least one relevant memory in the top 15. recall@15 is relevant records found over relevant records.

MindSpan categoryQACorrecthit@15
Family life85.0%51/6088.3%53/60
Craft mastery100.0%41/41100.0%41/41
Episodic precision94.7%36/3894.7%36/38
Career decisions100.0%33/33100.0%33/33
Social dynamics96.8%30/31100.0%31/31
Early childhood93.3%28/3096.7%29/30
Daily living100.0%29/29100.0%29/29
Home maintenance88.9%24/2792.6%25/27
Eldercare logistics95.5%21/2295.5%21/22
Health wellness95.5%21/22100.0%22/22
Logistical grounding94.1%16/17100.0%17/17
Kinship heritage87.5%14/1681.3%13/16
Enterprise architecture100.0%12/12100.0%12/12
Temporal anchors100.0%12/1291.7%11/12
Other (3 small categories)80.0%8/1090.0%9/10
Overall94.0%376/40095.5%382/400

QA is gpt-4o-mini answers, gpt-4o-mini LLM judge, temperature 0, k=15. Extraction: gemini-3.1-flash-lite. Embeddings: gemini-embedding-001. Dev 189/200 (94.5%), heldout 187/200 (93.5%). The hit@15 column is retrieval, not accuracy. Categories with fewer than 12 questions are Community engagement (1/2), Financial decisions (2/2), and Biographical chronology (5/6), folded into Other (8/10).

LoCoMo categoryQACorrect
1 · Multi-hop73.8%208/282
2 · Temporal69.8%224/321
3 · Open-domain51.0%49/96
4 · Single-hop90.7%763/841

Evidence recall is retrieval, not QA accuracy, on 1,535 LoCoMo questions that have evidence.

k@1@5@10@15@20@50
Evidence recall45.1%66.6%73.8%77.9%81.0%88.9%

Methodology

Each published number stays next to the protocol that produced it.

MindSpan, 94.0%
gpt-4o-mini answers, gpt-4o-mini LLM judge, temperature 0, k=15. Extraction: gemini-3.1-flash-lite. Embeddings: gemini-embedding-001. Dev 189/200 (94.5%), heldout 187/200 (93.5%).
Combined 376/400. Dev 189/200 (94.5%). Heldout 187/200 (93.5%).
LoCoMo, 80.8% (1,244/1,540)
gpt-4o-mini answers, gpt-4o-mini LLM judge, temperature 0. Extraction: gemini-3.1-flash-lite. Embeddings: gemini-embedding-001. LoCoMo standard protocol. Neutral default soul, not a dataset persona.
95% CI 78.7–82.7%, one run. Categories 1–4 only. Adversarial category 5 is excluded. Neutral default soul, not a dataset persona.
MindSpan hit@15 95.5% and recall@15 89.2%
382/400 questions have at least one relevant memory in the top 15. 463/519 relevant records. Retrieval, not answer accuracy. Same extraction and embedding models as the MindSpan QA figure.
LongMemEval keyless, session hit@10 96.8%, turn hit@10 87.2%
LongMemEval-S (cleaned), 100-question sample, 94 questions with labeled evidence scored. Fully local: no LLM anywhere. Embeddings: all-MiniLM-L6-v2 (ONNX, 384-d). Entity extraction: GLiNER small v2.1 (ONNX). Disk-backed store, reopened before recall. Retrieval only. Labeled evidence in the top k. Not answer accuracy. Zero LLM calls and zero API keys.
LoCoMo evidence recall@k
Share of gold evidence present in the top k memories. 1,535 questions have evidence. This is not QA accuracy and not hit@k. No LLM judge.

FAQ

Can I download ebbmem?

No. Use the hosted HTTP API or the hosted MCP endpoint.

Where is the public hostname?

It is not on this site. The paths above are relative to the origin you are given with access.

Where do I get a credential?

This site does not issue keys. The API documents register and login. Talk to us if you need access. Email support@ebbmem.ai.

What do memory limits refer to?

Memory limits refer to text memories (facts, messages and notes, typically a few KB each). Large files such as audio, video or images aren't stored as memories. Contact us for media and attachment storage options.