Memory & context infrastructure for AI agents

Persistent memory for AI agents. Lower context cost. Runtime control.

Belcore sits in front of your LLM to retain important context, reduce repeated token usage, and add runtime control.

This engine runs in production — a consumer application live on Google Play, with the iOS build in App Store review. It was not built for a benchmark and then pointed at users: it came out of real traffic, and the numbers below were measured afterward.

Get API Key — Free Try the live demo

1,000 calls/month · No card required

Projects

API Keys

Usage

Usage trendread-only

Model-agnostic · OpenAI · Gemini · any chat stack

96.79%

Memory retrieval · Recall

Gold coverage on LongMemEval_S — the retrieved context contains the answer-bearing turn.

It remembers across sessions

Across 500 multi-session conversations, the context Belcore assembles contains the turn that actually answers the question — the precondition for staying consistent over weeks.

This is retrieval recall (gold coverage): 96.79% overall, 95.95% multi-session. It measures whether the assembled context contains the answer-bearing turn — a retrieval metric, not end-to-end task accuracy.

93.88% ↓

Without re-sending the transcript

State is recomputed every turn, so history does not have to be replayed in full. You pay for signal, not for the whole conversation on every call.

6,671 input tokens per call vs 109,079 uncut, measured on LongMemEval_S with the o200k_base encoding.

3 modules

And it holds under attack

Injection attempts are caught server-side and deterministically — never delegated to the model. The agent deflects in persona, so an attack does not surface as a visible failure.

Provisional patent applications filed. These are filings, not granted patents.

Benchmark: LongMemEval_S, 500 questions, ~109k tokens of prior conversation each, fixed seed and frozen harness. Other memory vendors publish token and accuracy figures on different datasets and metrics, so we state our own scope rather than a side-by-side we cannot make apples-to-apples. Emotional continuity has no shared industry benchmark — we make no comparative claim for it.

The platform

Memory, cost, and control — in one layer.

Three surfaces for teams shipping AI agents to production. Each is backed by the real Belcore product below — not a mockup.

01 · Long-term memory

Retain what matters across sessions.

No sign-up, no cost. Move the sliders for recency and retention weighting, and the engine recomputes the retained state live — the same state it carries forward so an agent stays consistent across turns and sessions.

Input — recency, retention, context depth

Time since last interaction36 h

Retention weight70

Volatility45

Decay resistance25

Context depth8 Lv

Output — state computed by the engine

Context salience0

Drift signal0

Prompt directive

Idle for now. Try raising the time since last interaction and retention weight — once the values cross the threshold, the engine assembles a runtime directive and applies it here.

Telemetry

State transition executed
Context continuity maintained

Tokens avoided: 0
Latency: — ms
LLM calls: 0

A free preview that runs on pure computation, no LLM. The demo returns only output and telemetry — the directive wording, scoring formula, and internal variable relationships stay server-side. Verified cost and consistency numbers come from a production pilot against your own traffic.

02 · Reduce context cost

Pay for signal, not the whole transcript.

Belcore recomputes state each turn instead of replaying history, so you stop re-sending the full transcript on every call. Sign in to the usage dashboard for calls, tokens, and the full billing breakdown — baseline vs. actual cost, verified savings, the Belcore fee, and your net.

03 · Belcore Control

Preview · Coming soon

Enterprise-grade runtime control.

Track how an agent moves from data access to credentials, permissions, and external connections — and flag control deviations across the full behavior trajectory.

agentsupport-router-7f

Control deviation flagged

customer_db_readOn path
Customer data read
Within scope for this agent
network_accessOn path
Network access
Expected outbound lookup
credential_accessDeviation
Credential access
Not on the expected path after a data read
permission_changeDeviation
Permission change
Privilege escalation mid-trajectory
external_connectionBlocked
External connection
Blocked — credentials leaving the boundary

Trajectory scored across five capability types — data, network, credentials, permissions, and external egress. The path breaks after network_access, where credential and permission activity fall outside the expected sequence.

Get started

Start with a free key.

Create an account and your first key is issued instantly — no card required. Your keys, usage, and billing live behind your login, so you never paste a key to see your own account.

Create free account Sign in

Docs

Integrate in 10 seconds.

Put your key in the header and POST to /api/v1/chat. Send the session state along, and the updated values come back with the response (stateless).

Request · cURL

curl -X POST "https://belcore.xyz/api/v1/chat" \
  -H "Authorization: Bearer sk_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o-mini",
    "system": "You are the support assistant for Acme. Answer from the customer order history and keep replies concise.",
    "messages": [{ "role": "user", "content": "where is my order?" }],
    "state":   { "salience": 95, "drift": 20 },
    "weights": { "retention": 70, "volatility": 50, "stability": 60 },
    "contextDepth": 8
  }'

Response · JSON

{
  "reply": "Your order #1043 shipped yesterday and should arrive today.",
  "state": { "salience": 60, "drift": 12 },  // auto-decays after the exchange
  "directiveActive": true,   // a runtime directive was applied — the text stays server-side
  "model": "openai/gpt-4o-mini",
  "usage": { "inputTokens": 180, "outputTokens": 44 },
  "degraded": false          // true if Belcore fell back to plain passthrough
}

Endpoints

POST /api/v1/chatChat with the memory & context engine attached (core product)
POST /api/v1/emotion/advanceCompute the context-state transition only (no LLM · low cost)
GET /api/v1/accountUsage + billing breakdown across company, projects, and keys
GET /api/v1/account/invoiceMonthly invoice — Belcore 10% of verified savings
GET /api/v1/usageCheck this month's call count
GET /api/v1/safetyLive reliability posture — retention, failover, circuit breaker, isolation

Pricing

Start free. Scale as your savings grow.

Free

Instant API key, no card required
1,000 chat calls per month
Emotion-state endpoint (no LLM), fair use
Live demo and docs

Get API key

Growth

from$99/ month

$99/month minimum, or 10% of verified savings — whichever is greater
Full API access, projects, keys, and usage dashboard
Isolation by company, project, and API key
Email support

Start with Growth

Enterprise

Custom

Annual contract with SLA and dedicated support
Security review pack: isolation evidence, audit trail, and retention terms
Belcore Control add-on (Preview, coming soon)
Custom retention and deployment terms

Talk to us

How verified savings works

We establish a baseline using your pre-Belcore token usage or an agreed benchmark period, then compare it with your actual token cost after Belcore is enabled. If 10% of verified savings is below the monthly minimum, the minimum applies.

Prices exclude applicable taxes.

Contact us

From billing, accounts, and service errors to API integration — pick a category and we'll collect only what we need. We usually reply to the email you provide within one business day.

Measured evidence