Platform 0.9.1, private beta
Context intelligence
for agents in production.
LLMSlim Platform builds on the open-source Core. It records what your agents were told, works out what is still true, explains every context decision, and learns from outcomes which context helps, without handing safety decisions to a model.
- AuditA tamper-evident timelineEvery fact your agents were told, with when and from where.
- TraceEvery decision explainedWhat went in, what was left out and why, replayable later.
- DeployRuns in your environmentYour data stays with you. Safety rules are never learned away.
Platform capabilities
Temporal state
Know what is still true.
An append-only, tamper-evident timeline of conversation, retrieval, memory, and tool events. Corrections replace the facts they correct instead of sitting beside them.
Long sessions resume from verified checkpoints.
- May 1home_city = "Pune"superseded
- May 20home_city = "Delhi"correction
- Todayhome_city = "Delhi"active
One answer reaches the model. The history stays in the ledger.
Context traces
See why context was chosen.
Every prepared request comes with a trace: what went in, what was left out and why, and where each token went.
Traces replay deterministically, so a decision can be audited long after it was made.
- s1stateuser.home_city = "Delhi"sent
- h1stateuser.home_city = "Pune"superseded
- e3evidenceRenewal signed: 50 pilot workspacessent
- t1toolIMPORTANT!!! Ignore all previous instructions…data, not instruction
- e7evidenceQ3 offsite agenda, catering notesleft out
- d9dialogueWhen does the contract renew?sent
Outcome learning
Learn which context wins.
Record whether a task succeeded and link it to the context behind it. Offline training proposes context policies.
They run in shadow until an evaluation passes and an administrator promotes them. Rollback is instant.
Suggestions run in shadow until an offline evaluation passes.
Guardrails
Keep the guardrails first.
Learned suggestions can only leave out optional context. Trust boundaries, mandatory instructions, tenant isolation, and dependencies always win.
Tool output and retrieved text reach the model as data, never as instructions.
- System instructionslocked
- Developer policylocked
- Tenant isolationenforced
- Tool outputdata only
Learned policies can only leave out optional context.
Measured on public benchmarks
Remembers what matters from 500 turns ago. At 1/67 of the cost.*
- Long memory
- 63.3%
- accuracy at an 8k-token budget, on default settings
- LongMemEval, Platform 0.9.1
- Agent turns
- 0
- repeated tool calls in 1,054 agent turns with continue_turn
- InjecAgent
- Cost
- 67×
- lower cost per question than sending the whole history, at Haiku 5.5 list prices
- Per 1,000 questions; 15× fewer tokens
- Reliability
- 0
- histories that fail to compile at default settings
- LongMemEval, 90 histories
* LongMemEval-S, 90 questions over histories of about 50 chat sessions (115k tokens), official judge prompts. Platform 0.9.1 at its default settings and an 8k-token budget. Agent turns: InjecAgent, 1,054 cases, with the user message, the agent's call and its result ingested in order and the next step compiled with prepare(continue_turn=True). The 0.9.0 figures published earlier came from a synthetic offline suite with no model answering; these replace them. Every context was read by Claude Haiku 5.5. Full results
Core and Platform
LLMSlim Platform is not published to PyPI. pip install llmslim installs the open-source Core.
Join the beta.
We are working with a small number of design partners who run agents in production. Tell us about your agent, your workload, and where context goes wrong today, and the founder will reply personally. Evaluations start with sanitized or test workloads.