lerim 0.1.66

Continual learning layer for coding agents and software projects.

Your coding agents forget everything after each session.
Lerim learns — across all of them.

lerim.dev · docs · pypi

The Problem

You spend 20 minutes explaining context to your coding agent. It writes great code. Next session? It's forgotten everything. Every decision, every pattern, every "we tried X and it didn't work" -- gone.

And if you use multiple agents -- Claude Code at the terminal, Cursor in the IDE, Codex for reviews -- none of them know what the others learned. Your project knowledge is scattered across isolated sessions with no shared intelligence.

This is agent context amnesia, and it's the biggest productivity drain in AI-assisted development.

The Solution

Lerim is the continual learning and context graph layer for AI coding agents — it watches sessions, extracts structured knowledge, and builds a shared intelligence graph across agents, projects, and teams.

• Watches your agent sessions across all supported coding agents
• Extracts decisions and learnings automatically using LLM pipelines (DSPy + PydanticAI)
• Stores everything as plain markdown files in your repo (.lerim/)
• Refines knowledge continuously -- merges duplicates, archives stale entries, applies time-based decay
• Connects learnings into a context graph -- related decisions and patterns are linked
• Unifies knowledge across all your agents -- what one agent learns, every other can recall
• Answers questions about past context: lerim ask "why did we choose Postgres?"

No proprietary format. No database lock-in. Just markdown files that both humans and agents can read. Knowledge compounds over time, not stale.

Supported Agents

Lerim works with any coding agent that produces session traces. Current adapters:

Agent	Session Format	Status
Claude Code	JSONL traces	Supported
Codex CLI	JSONL traces	Supported
Cursor	SQLite to JSONL	Supported
OpenCode	SQLite to JSONL	Supported

Adding a new agent adapter is straightforward -- PRs welcome! See src/lerim/adapters/ for examples.

How It Works

Lerim is file-first and primitive-first.

• Primitive folders: decisions, learnings, summaries
• Project memory: <repo>/.lerim/
• Global fallback: ~/.lerim/
• Search: file-based (no index required)
• Orchestration: pydantic-ai lead agent + read-only explorer subagent
• Extraction/summarization: dspy.ChainOfThought with transcript windowing

Sync path

The sync path processes new agent sessions: reads transcript archives, extracts decision and learning candidates via DSPy, deduplicates against existing knowledge, and writes new entries to the project's knowledge store.

Maintain path

The maintain path runs offline refinement over stored knowledge: merges duplicates, archives low-value entries, consolidates related learnings, and applies time-based decay to keep the context graph clean and relevant.

Quick start

1. Install

pip install lerim

Prerequisites: Python 3.10+, Docker (optional)

2. Set up

lerim init                     # interactive setup — detects your coding agents
lerim project add .            # add current project (repeat for other repos)

3. Set API keys

Set keys for the providers you configure (defaults: MiniMax primary, Z.AI fallback):

export MINIMAX_API_KEY="..."   # if using MiniMax (default)
export ZAI_API_KEY="..."       # if using Z.AI (default fallback)

You only need keys for providers referenced in your [roles.*] config. See model roles.

4. Start Lerim

lerim up

That's it. Lerim is now running as a Docker service — syncing sessions, extracting decisions and learnings, refining memories, and serving a dashboard at http://localhost:8765.

5. Teach your agent about Lerim

Install the Lerim skill so your agent knows how to query past context:

lerim skill install

This copies skill files (SKILL.md, cli-reference.md) into your agent's skill directory.

6. Get the most out of Lerim

At the start of a session, tell your agent:

Check lerim for any relevant memories about [topic you're working on].

Your agent will run lerim ask or lerim memory search to pull in past decisions and learnings before it starts working.

Running without Docker

If you prefer not to use Docker, Lerim works directly:

lerim connect auto             # detect agent platforms
lerim serve                    # run HTTP server + daemon loop

Local models (Ollama)

Lerim works with local models via Ollama. Set provider = "ollama" in your role config and Lerim will automatically load models into GPU/RAM before each sync/maintain cycle and unload them after to free memory (auto_unload = true in [providers]).

ollama serve                   # start Ollama (runs outside Docker)

For Docker deployments, set ollama = "http://host.docker.internal:11434" in [providers] so the container can reach the host Ollama instance. See model roles for full configuration.

Dashboard

The dashboard gives you a local UI for session analytics, knowledge browsing, and runtime status.

Tabs

• Overview: high-level metrics and charts (sessions, messages, tools, errors, tokens, activity by day/hour, model usage).
• Runs: searchable session list with status and metadata; open any run in a full-screen chat viewer.
• Memories: library + editor for memory records (filter, inspect, edit title/body/kind/confidence/tags).
• Pipeline: sync/maintain status, extraction queue state, and latest extraction report.
• Settings: dashboard-editable config for server, model roles, and tracing; saves to ~/.lerim/config.toml.

CLI reference

Full command reference: skills/lerim/cli-reference.md

# Setup (host-only)
lerim init                                  # interactive setup wizard
lerim project add ~/codes/my-app            # register a project
lerim project list                          # list registered projects

# Docker service
lerim up                                    # start Lerim container
lerim down                                  # stop it
lerim logs --follow                         # tail logs

# Alternative: run directly without Docker
lerim serve                                 # start HTTP server + daemon loop

# Service commands (require lerim up or lerim serve running)
lerim ask "Why did we choose this?"          # query memories
lerim sync                                  # one-shot: sync sessions + extract
lerim maintain                              # one-shot: merge, archive, decay
lerim status                                # runtime state

# Local commands (run on host, no server needed)
lerim memory search "auth pattern"          # keyword search
lerim memory list                           # list all memories
lerim memory add --title "..." --body "..." # manual memory
lerim connect auto                          # detect and connect platforms
lerim skill install                         # install skill into agent directories

Configuration

TOML-layered config (low to high priority):

1. src/lerim/config/default.toml (shipped with package -- all defaults)
2. ~/.lerim/config.toml (user global)
3. <repo>/.lerim/config.toml (project overrides)
4. LERIM_CONFIG env var path (explicit override, for CI/tests)

API keys come from environment variables only. Set keys for the providers you use:

Variable	Provider	Default role
MINIMAX_API_KEY	MiniMax	Primary (all roles)
ZAI_API_KEY	Z.AI	Fallback
OPENROUTER_API_KEY	OpenRouter	Optional alternative
OPENAI_API_KEY	OpenAI	Optional alternative
ANTHROPIC_API_KEY	Anthropic	Optional alternative

Default model config (from src/lerim/config/default.toml):

• All roles: provider=minimax, model=MiniMax-M2.5
• Fallback: zai:glm-4.7 (lead/explorer), zai:glm-4.5-air (extract/summarize)

Development

uv venv && source .venv/bin/activate
uv pip install -e '.[test]'
lerim init                    # first-time config
lerim project add .           # track this repo
lerim up                      # start the service
tests/run_tests.sh unit
tests/run_tests.sh all

Tracing (OpenTelemetry)

Lerim uses PydanticAI's built-in OpenTelemetry instrumentation for agent observability.

# Enable tracing
LERIM_TRACING=1 lerim sync

# or in config
# .lerim/config.toml
[tracing]
enabled = true

View traces at https://logfire.pydantic.dev.

Memory layout

Project scope:

<repo>/.lerim/
  config.toml              # project overrides
  memory/
    decisions/
    learnings/
    summaries/
      YYYYMMDD/
        HHMMSS/
          {slug}.md
    archived/
      decisions/
      learnings/
  workspace/
    sync-<YYYYMMDD-HHMMSS>-<shortid>/
    maintain-<YYYYMMDD-HHMMSS>-<shortid>/
  index/   # reserved

Global fallback scope follows the same layout under ~/.lerim/.

Contributing

Lerim is open to contributions. Whether it's a new agent adapter, a bug fix, or a documentation improvement, PRs are welcome.

• Read the Contributing Guide
• Browse open issues
• Agent adapter PRs are especially appreciated -- see src/lerim/adapters/ for examples

Docs

Full documentation: docs.lerim.dev

• Quickstart
• Installation
• CLI Reference
• Configuration
• Architecture
• Contributing

---

If Lerim saves you from re-explaining context to your agent, give it a ⭐
Star on GitHub