Custom & Non-Coding Agents

Start here if your agent is not a coding agent with a native Lerim adapter. That covers support, incident/security operations, research, compliance, revenue, and any custom business agent that produces completed runs but does not live in an IDE session store.

Coding agents (Claude Code, Codex CLI, Cursor, OpenCode, pi) have native adapters that read completed sessions automatically. Everything else uses the custom agent path below. Both paths feed the same compiler and the same context store.

1. Does your agent have a native adapter?

If your agent runs as a coding agent in a local session store, use the native adapter path instead — it is automatic:

lerim connect auto
lerim project add .
lerim up

See Connecting Agents for the adapter list. If your agent is API-driven, server-side, or part of a business workflow (support, incidents, research, compliance, sales, internal tools), continue below.

2. Pick or write a signal profile

A signal profile tells Lerim's extraction what counts as durable signal for your workflow and what is noise. Lerim bundles six profiles:

Profile	Use for
coding	Repository and coding-agent work (default).
support	Customer support and customer operations.
ops	Incident response, operations, and reliability.
research	Research, market intelligence, and analysis.
compliance	Compliance, legal, regulatory, and policy review.
generic	General-purpose fallback when no other profile fits.

List them and inspect one:

lerim profile list
lerim profile show research

If none fit — for example a sales, procurement, or internal business agent — write your own YAML profile. It is four rule groups (focus_rules, reject_as_noise, evidence_rules, scope_rules) and takes a few minutes. See Customize Lerim For Your Use Case.

3. Get a clean completed-run trace

Lerim reads completed sessions, not live in-flight state. Export your agent's finished run as canonical JSONL, one event per line:

{"type":"user","message":{"role":"user","content":"..."},"timestamp":"2026-05-16T09:00:00Z"}
{"type":"assistant","message":{"role":"assistant","content":"..."},"timestamp":"2026-05-16T09:02:00Z"}

JSON arrays, wrapper objects (messages / events / trace), and plain text are also accepted. See Submit A Custom Agent Trace for the full shape reference.

Clean before import. Remove secrets, tokens, regulated personal data, and huge raw payloads before the trace enters Lerim. Lerim extraction is selective, but it is not a privacy firewall. For a quick first pass, use the defense-in-depth redaction helper.

4. Import it

One file:

lerim trace import ./my-agent-run.jsonl \
  --source-name support-agent \
  --source-profile support \
  --scope-type domain \
  --scope support-ops

Or register an ongoing folder of clean traces so background ingest picks up new files automatically:

lerim project add ~/lerim-traces/support-clean --type custom --source-profile support
lerim ingest --agent custom

See Custom Trace Folders.

5. See the extracted context

lerim context records --source-profile support
lerim status

Routine sessions can produce zero durable records — that is the compiler rejecting noise, not a failure. A useful run produces a few cited records across decision, constraint, fact, preference, and episode.

6. Query it for the next run

lerim answer "What approval constraints should the next support agent know?"
lerim context-brief show

For MCP-capable agents, install Lerim as an MCP server so the agent can query context in-loop: MCP Quickstart.

7. The improvement loop

Approved traces, corrections, and extracted records are structured, cited, and scoped — which makes them reusable beyond the next query. They become training-ready data for the same workflow (for example, distilling a strong general model into a cheaper task-specific one).

Lerim has two layers. The open core (Apache-2.0, this repo) captures, compiles, cites, and serves context. Model specialization — distillation, RL, and prompt and harness tuning for a specific workflow — is the private Lerim layer, built on top of this open foundation.

Worked examples

Each of these shows a realistic trace, the real records Lerim extracted, and what the next agent now knows:

• Support Ops Demo
• Incident Ops Demo
• Research Demo
• Compliance Demo