Stop restarting from zero.
Carry forward decisions, constraints, failed paths, and handoffs without replaying every transcript.
Apache-2.0 context compiler for AI agent teams
Agents should not rediscover yesterday's decisions.
Lerim turns completed agent sessions into cited context records: decisions, constraints, facts, handoffs, and evidence future agents can retrieve before they work.
Self-host the Apache-2.0 core. Hire Lerim for agent harness engineering, prompt and context optimization, private deployment, and SFT/RL model specialization.
Capture
Completed source sessions from CLI, MCP, or JSONL
Compile
Decisions, constraints, facts, handoffs, and evidence
Reuse
Scoped answers and startup briefs for future agents
Open core
CLI, adapters, MCP server, context DB, and local runtime
Enterprise help
Harness engineering, prompt evals, private deployment, SFT/RL training, and support
First step
A focused Context Audit on one repeated workflow
Why it matters
Make agent work survive the run.
Raw traces are too long, too noisy, and too platform-specific to replay directly. Lerim compiles the small operating context future agents actually need.
Keep source evidence attached.
Context records stay linked to the session that produced them, so later agents can inspect why the memory exists.
Make memory selective.
Routine chatter, temporary steps, and source-derivable facts should not become permanent operating context.
Enterprise services
Apache-2.0 gives teams the core. We make the agent harness production-grade.
Teams can self-host Lerim. We sell the production work around it: harness design, prompt/context optimization, eval loops, private deployment, and SFT/RL model specialization.
Lerim Trace Integration
Connect customer agent runs, tools, repos, tickets, chats, and business workflows into Lerim’s trace ingestion layer. Build clean adapters, redaction, scopes, and source-session capture so Lerim can compile reliable context.
Context Compiler Optimization
Tune Lerim’s compiler for the customer’s domain: what becomes memory, what gets rejected, how records are merged, cited, retrieved, and injected into future agent runs.
Private Lerim Deployment
Deploy Lerim as the customer’s private context layer: MCP server, context database, eval workflows, audit logs, retention, monitoring, SSO/RBAC, and production support.
Lerim-Native Model Training
Train private models inside the Lerim harness using approved traces, context records, corrections, and evals so repeated workflows run cheaper, faster, and with customer-specific context.
First paid step
Start with a Context Audit.
Pick one repeated AI-agent workflow. We inspect the current harness, trace quality, prompts, context flow, and evals, then deliver a concrete optimization and training roadmap.
Where it fits
Reusable context for repeated business workflows.
The best first deployment is a repeated, evidence-heavy workflow where agents rediscover decisions, constraints, or handoffs between runs.
Research and market intelligence
Source trails, evidence strength, assumptions, rejected leads, brief constraints, and analyst handoffs.
Support and success
Customer constraints, failed fixes, product behavior, escalation reasons, policy references, and next steps.
Security and incidents
Investigation timelines, owner decisions, policy exceptions, remediation evidence, runbook lessons, and unresolved risks.
Engineering automation
Repo conventions, architecture decisions, failed tests, release lessons, and implementation handoffs.
How it works
A compiler pass between completed runs and future agents.
Lerim is not a broad manual memory bucket. Completed sessions become compact, cited context only after the compiler decides the signal is reusable.
01
Connect source sessions
Read native coding-agent histories or import clean custom traces from business workflows.
02
Compile durable context
Extract reusable decisions, constraints, facts, preferences, handoffs, and evidence while rejecting weak signal.
03
Deliver
Serve context into the agent harness through CLI answers, startup briefs, MCP tools, and hosted or private deployment surfaces.
Integration shape
Native capture where possible. MCP everywhere else.
Native adapters are best when an agent has a stable local session store. MCP gives compatible clients recall and completed-session submission without pretending every client has native capture.
Current foundation
Local CLI/runtime, SQLite context store, BAML/LangGraph extraction, native coding-agent adapters, generic trace import, MCP server, context briefs, and cited answers.
Commercial path
Agent harness engineering, prompt/context optimization, private MCP, SSO/RBAC, audit logs, retention, and specialized SFT/RL training sit above the Apache-2.0 core.
What a pilot proves
Evidence first. No inflated memory claims.
A serious rollout should prove the actual workflow: does Lerim find useful evidence, reduce repeated context work, and avoid filling memory with noise?
Find the right source run.
Measure whether future agents retrieve the session that contains the decision, constraint, or evidence they need.
Use less context, not less truth.
Track how much transcript text is avoided while preserving the source evidence needed for the next run.
Reject low-value memory.
Review false positives, duplicates, vague notes, and source-derivable facts before scaling the workflow.
Start here
Bring one repeated workflow.
We will connect the traces, tune the compiler, and prove whether Lerim gives future agents better startup context than raw transcripts or manual notes.
Good audit candidates
- Agents keep rediscovering decisions, constraints, policies, or failed paths
- Teams paste old transcripts into new runs because context is not reusable
- Security or platform teams need private deployment, audit, and retention controls