Evolve: On‑the‑job learning for AI agents

Coding agents repeat the same mistakes because they start fresh every session. Evolve gives agents memory — they learn from what worked and what didn't, so each session is better than the last.

Evolve is a system designed to help agents improve over time by learning from their trajectories. The Lite version is designed to effortlessly slot into existing agent assistants like Claude Code and Codex. It uses a combination of an MCP server for tool integration, vector storage for memory, and LLM-based conflict resolution to refine its knowledge base.

On the AppWorld benchmark, Evolve improved agent reliability by +8.9 points overall, with a 74% relative increase on hard multi-step tasks. Evolve is a system designed to help agents improve over time by learning from their trajectories. It uses a combination of an MCP server for tool integration, vector storage for memory, and LLM-based conflict resolution to refine its knowledge base.

Important

⭐ Star the repo: it helps others discover it.

Latest from Evolve

Research, releases, and practical findings from the Evolve team.

• 2026-09-25 · Paper — Memory as Middleware for Self-Improving AI Agents The case for agent memory as a pluggable middleware layer, with Evolve as the reference implementation (to appear at Middleware 2026).
• 2026-09-15 · Article — Your Agent Aced the Task. Will It Do It Again? Diagnosing where an agent's decisions wobble and stabilizing them so success repeats.
• 2026-09-08 · Paper — Closing the Consistency Gap: Self-Evolving Agents That Learn to Stay on Course Turning unstable trajectory steps into episodic memory to keep agents consistent across repeated runs.
• 2026-08-18 · Article — How Much Memory Does Your Agent Actually Need? Why agent memory should be calibrated to the model instead of simply increased.
• 2026-08-11 · Article — Thinking of ACE? We Can Do It with Fewer Tokens Comparing Evolve's targeted guideline retrieval with ACE's playbook approach and token usage.
• 2026-04-08 · Article — ALTK‑Evolve: On‑the‑Job Learning for AI Agents An introduction to learning reusable guidance from agent trajectories.
• 2026-04-07 · Announcement — IBM introduces ALTK Evolve Evolve is opened to builders creating agents that learn from experience.
• 2026-03-11 · Paper — Trajectory-Informed Memory Generation for Self-Improving Agent Systems The research framework and evaluation underlying Evolve.

Quick Start (Lite)

IBM Bob →

Claude Code →

Codex →

Hermes →

Quick Start (Evolve MCP Server)

Installation

Prerequisites:

• Python 3.12 or higher
• uv (recommended) or pip

From Source

# Clone the repository and install dependencies
git clone https://github.com/agenttoolkit/altk-evolve.git
cd altk-evolve
uv venv --python=3.12 && source .venv/bin/activate
uv sync
# Build the UI
cd frontend/ui
npm ci && npm run build
cd ../..

From PyPI

pip install altk-evolve

Optional Backend Dependencies:

The default filesystem backend uses simple text matching and requires no additional dependencies. For semantic vector similarity search, install one of these backends:

For PostgreSQL with pgvector support (recommended for production):

uv sync --extra pgvector

For Milvus support (optimized for large-scale vector search):

uv sync --extra milvus

See the Backend Configuration Guide for detailed comparison and setup instructions.

Configuration

For direct OpenAI usage:

export OPENAI_API_KEY=sk-...

For LiteLLM proxy usage and model selection (including global fallback via EVOLVE_MODEL_NAME), see the configuration guide.

Running Services

Start the Web UI and MCP server

uv run evolve-mcp

The Web UI can be accessed from: http://127.0.0.1:8000/ui/

Starting the Web UI and MCP Server

If you only want to access the Web UI and API (without the MCP server stdio blocking the terminal), you can run the FastAPI application directly using uvicorn:

uv run uvicorn altk_evolve.frontend.mcp.mcp_server:app --host 127.0.0.1 --port 8000

Then navigate to http://127.0.0.1:8000/ui/.

Starting only the MCP Server

If you're attaching Evolve to an MCP client that requires a direct command (like Claude Desktop):

uv run evolve-mcp

Or for SSE transport:

uv run evolve-mcp --transport sse --port 8201

Verify it's running:

npx @modelcontextprotocol/inspector@latest http://127.0.0.1:8201/sse --cli --method tools/list

Available tools:

• get_entities(task: str, entity_type: str = "guideline", include_public: bool = False): Get relevant entities for a specific task. Set include_public=True to merge in public entities from all other namespaces; those results are annotated with [public: {owner_id}].
• get_guidelines(task: str): Get relevant guidelines for a specific task (backward compatibility alias for get_entities).
• get_guidelines_with_attribution(task: str): Return formatted guidelines with the entity IDs used to build the prompt.
• get_relevant_guidelines(task: str, top_k: int | None, core_support: int | None): Retrieve the always-on guideline core plus a task-relevant dosage.
• list_entities(...): Return structured, filtered, cursor-paginated entity inventory for user and administrative UIs.
• get_entity(entity_id: str, user_id: str | None, record_access: bool = True): Return one structured entity, enforcing ownership when a caller ID is supplied.
• patch_entity_metadata(entity_id: str, metadata_patch: str, user_id: str | None): Merge JSON metadata through the memory hook seam.
• record_access(entity_ids: list[str], accessed_at: str | None): Explicitly stamp the retention engine's last_accessed signal.
• validate_retention_policy(policy: str): Validate and normalize a JSON retention policy without scanning data.
• put_retention_policy(policy_id: str, name: str, policy: str, ...): Create or replace an Evolve-owned retention policy.
• get_retention_policy(policy_id: str) / list_retention_policies(): Read the namespace's policy catalog for operators and management UIs.
• run_retention(policy_id: str, dry_run: bool = True, ...): Dry-run or apply a stored policy and persist a structured, entity-linked report.
• list_retention_runs(...): Read namespace-scoped retention run history, optionally filtered by agent or policy.
• get_compliance_status(): Report backend health, retention availability, hook coverage, and configured plugin health.
• save_trajectory(trajectory_data: str, task_id: str | None, owner_id: str | None): Save a conversation trajectory and generate new guidelines.
• create_entity(content: str, entity_type: str, metadata: str | None, enable_conflict_resolution: bool, owner_id: str | None, visibility: str = "private"): Create a single entity. Pass visibility="public" and owner_id to make it immediately discoverable by other namespaces.
• publish_entity(entity_id: str, user_id: str | None): Make an entity publicly visible to all namespaces. Records the caller as owner and stamps published_at.
• unpublish_entity(entity_id: str, user_id: str | None = None): Revert an entity to private visibility. Ownership is enforced server-side: if the entity has an owner_id, user_id must match it.
• delete_entity(entity_id: str): Delete a specific entity by its ID.

Filter Migration Note

Entity search filters reserve bare keys for top-level schema columns only: id, type, content, and created_at.

If you need to filter on JSON metadata, use the metadata.<key> form. For example, use filters={"type": "trajectory", "metadata.task_id": "123"} instead of filters={"type": "trajectory", "task_id": "123"}.

Existing integrations that stored custom fields in entity metadata should update filter writers to add the metadata. prefix for those keys.

Features

• Proactive: Learns how to recognize problems and their solutions, and generates guidelines that get automatically applied to new tasks.
• Conflict Resolution: Update existing guidelines when new information contradicts them.
• On Command: An array of tools to manage guidelines whether in the agent or through a CLI
• Sharing: Publish individual entities so other agents can discover and retrieve them across namespaces.

Architecture

Evolve is built on a modular architecture which forms a feedback loop, taking conversation traces (trajectories) from an agent, extracting key insights into a database, feeding it back into the agent.

Lite Mode omits the Interaction layer. All activity is performed in-agent

Entity Sharing

Evolve supports sharing entities across namespaces using a simple public/private visibility model.

Visibility is stored in each entity's metadata and is private by default. Existing entities without a visibility field are unaffected.

Metadata field	Description
owner_id	User ID who created or last published the entity
visibility	"private" (default) or "public"
published_at	ISO-8601 timestamp of the most recent publish

MCP Tools

Personal facts on a shared Evolve service:

store_user_facts(namespace_id="service-instance-1", user_id="alice", message="I prefer concise answers")
retrieve_user_facts(namespace_id="service-instance-1", user_id="alice", query="answer preferences")

Pass the service instance ID as namespace_id and the individual user's ID as user_id on both calls. Explicitly scoped retrieval filters by that exact pair, including query fallback, and never falls back to another user's facts. Empty explicit namespaces or user IDs are rejected. Calls omitting namespace_id retain the configured default namespace and legacy default-user fallback. Integrating clients must supply the scope; upgrading Evolve alone cannot infer which service instance an unscoped request belongs to.

Publishing an entity:

publish_entity(entity_id="42", user_id="alice")

Sets visibility=public and records the owner and publish timestamp.

Unpublishing:

unpublish_entity(entity_id="42", user_id="alice")

Reverts the entity to private. The entity stays in its namespace — only its visibility changes.

Retrieving public entities from all namespaces:

get_entities(task="write safer code", include_public=True)

Merges results from the caller's namespace with public entities from all other namespaces. Public results are annotated with [public: {owner_id}].

Creating an entity with visibility:

create_entity(content="...", entity_type="guideline", visibility="public", owner_id="alice")

What's deferred (Phase 1C)

REST API endpoints (GET /api/entities/public, publish/unpublish routes) and UI controls are not yet implemented.

Guideline Provenance

Evolve automatically tracks the origin of every guideline it generates or stores. Every guideline entity contains metadata identifying its source:

• creation_mode: Identifies how the guideline was created (auto-phoenix via trace observability, auto-mcp via trajectory saving tools, or manual).
• source_task_id: The ID of the original trace or task that inspired the guideline, providing full auditability.

See the Low-Code Tracing Guide for more details.

Contributing, Community, and Feedback

Evolve is an active project, and real‑world usage helps guide its direction.

If you’re experimenting with Evolve or exploring on‑the‑job learning for agents, feel free to open an issue or discussion to share use cases, ideas, or feedback.

See the Contributing Guide to understand our development process, or how to submit changes, report bugs, or propose features.

Embedded memory API and scheduling

Hosts can mount the scoped REST router with their own client and authentication dependencies. Evolve owns retention schedules and worker execution, using Kubernetes-compatible cron, timezone, concurrency, deadline, and suspension fields. Use evolve retention schedules to manage schedules and start/stop to enable or suspend them. The Evolve service owns background execution.

The public retention service is available as client.retention(namespace_id, agent_id=...); CLI, REST, MCP, and scheduling share its operations and scope checks.

OpenTelemetry service instrumentation

Evolve's Python operations use the OpenTelemetry API and inherit the host's provider. Constructing EvolveClient never installs or replaces a provider. Standalone MCP services export telemetry by default. Compatible SDK and exporter versions are declared as dependencies and enforced at installation. To configure a destination:

uv pip install altk-evolve
export OTEL_SERVICE_NAME=evolve
export OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4318
export OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf
python -m altk_evolve.frontend.mcp --transport sse

Without an endpoint setting, the OTLP exporter uses localhost:4318. Export runs in the background with bounded queues; an absent collector does not block memory operations. Unsent telemetry may be dropped. Requests default to a one-second export timeout, overridable through standard OTEL timeout settings. Set EVOLVE_OTEL_ENABLED=false to disable Evolve-owned export. If exporter packages cannot be imported, Evolve logs a warning and continues without an exporter. Invalid exporter configuration also logs a warning and leaves service startup unaffected; correct the settings and retry initialization to enable export.

Use an actual reachable collector address. OpenLIT and other OTLP backends are supported without their SDKs. This exporter uses HTTP/protobuf; the signal-specific OTEL_EXPORTER_OTLP_TRACES_ENDPOINT and OTEL_EXPORTER_OTLP_METRICS_ENDPOINT include /v1/traces and /v1/metrics, respectively. Standard resource, sampling, batch-size, export-timeout and metric-interval environment settings apply. OTEL_TRACES_EXPORTER=none and OTEL_METRICS_EXPORTER=none disable each exporter; OTEL_SDK_DISABLED=true prevents Evolve's service initialization. An existing host tracer provider is left untouched, including ownership of its shutdown. Custom service hosts can explicitly call configure_service_telemetry() from altk_evolve.telemetry at startup. The SDK installs process-exit shutdown hooks; the MCP lifespan does not synchronously flush or wait for a collector.

Spans cover memory search, query embeddings and PostgreSQL retrieval, write preparation/commit, fact extraction and conflict resolution, model calls, processing stages/processors, active hooks, and retention marking/sweeping/jobs. FastMCP propagates MCP trace metadata; the scoped REST router accepts W3C trace headers. Independent scheduled attempts start separate traces with job references. Duration histograms and outcome counters have bounded operation/result labels, never namespace, user, conversation or entity IDs. Namespace attribution belongs on spans. Evolve's instrumentation excludes contents, prompts, SQL values and exception messages. Evolve-owned export forwards Evolve's instrumentation and sanitized FastMCP transport spans that retain timing, IDs, kind and status code to preserve the trace tree. Transport names are fixed; attributes, exception events, links and status descriptions are omitted. Metrics only forward the altk_evolve scope. A host-owned provider controls its own exporters and third-party capture settings. Traces are diagnostic and may be sampled; durable retention receipts remain the authoritative audit record.

MCP clients must send notifications/cancelled to cancel server work; stopping a local wait alone does not cancel the request. Evolve checks cooperative cancellation around model calls and before committing prepared memory writes. An in-flight provider call may still finish, and earlier committed writes remain. These checkpoints do not provide rollback or interrupt arbitrary synchronous work.