ACE Context Learning Engine

What does this agent do, and when should you use it?

Agentic Context Engine (ACE) is an open-source Python engine for adding a persistent learning loop to AI agents. It stores strategies in a persistent Skillbook and divides work among an Agent, Reflector, and SkillManager for execution, trace analysis, and strategy curation. Its Recursive Reflector writes and runs Python in a sandboxed environment to search traces for patterns, isolate errors, and derive actionable insights. The repository offers ACELiteLLM, ACE, TraceAnalyser, BrowserUse, LangChain, and ClaudeCode runners, plus a composable pipeline system. It can be used as a library in a Python project and exposes ace, kayba, and ace-mcp commands; Kayba is the managed service powered by the engine.

With ACELiteLLM, an application calls .ask() to invoke a model, then passes a correction to .learn_from_feedback(); ACE extracts a strategy, updates the Skillbook, and makes the strategy available to later .ask() calls. For existing logs, agent.learn_from_traces(your_existing_traces) provides a path to learn from recorded traces without rerunning tasks. A full learning flow can combine AgentStep, EvaluateStep, ReflectStep, UpdateStep, and DeduplicateStep, or use learning_tail() for the standard tail sequence. The environment sends a trace to the Reflector; the Recursive Reflector executes Python analysis in a sandbox, and the SkillManager adds, refines, or removes Skillbook strategies. The CLI supports ace setup for model and key configuration, ace models for model and pricing search, ace validate for connection testing, and kayba for trace upload, insight retrieval, and prompt management.

1. A Python developer whose question-answering or tool-using agent receives human corrections and needs those corrections to influence later requests.
2. A team with historical agent logs that wants to extract reusable strategies through learn_from_traces() instead of rerunning the original tasks.
3. A developer using browser-use for browser automation who wants repeated runs to improve through accumulated strategies.
4. A team already operating a LangChain chain or agent that wants to add a learning loop through the LangChain runner.
5. An engineering team running Claude Code CLI tasks that wants to apply the learning loop through the ClaudeCode runner.

How do you install or deploy this agent?

You need Python, uv, and an API key for a supported model provider. Install with:

uv add ace-framework

For guided configuration:

ace setup

Or set credentials manually; for OpenAI:

export OPENAI_API_KEY="your-key"

Install integration extras as needed:

uv add 'ace-framework[browser-use]'
uv add 'ace-framework[langchain]'
uv add 'ace-framework[mcp]'

How do you use this agent?

After running ace setup or setting an API key, a minimal learning flow is:

from ace import ACELiteLLM

agent = ACELiteLLM(model="gpt-4o-mini")
answer = agent.ask("Is there a seahorse emoji?")
agent.learn_from_feedback("There is no seahorse emoji in Unicode.")
answer = agent.ask("Is there a seahorse emoji?")
print(agent.get_strategies())

Use ace validate <model> to test a model connection, or ace models <query> to search available models and pricing.

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

1. v0.13.0 2026-09-23
   Release notes excerpt

   Fixes for current Claude models and for MCP Python SDK 2.x. Fixed - The default Claude model is now claude-sonnet-5. Anthropic retired the old default claude-sonnet-4-20250514 on 15 June 2026. (#141) - temperature is sent only when you set it. Claude 4.7 and later models return a 400 error for non-default values. (#141, #144) - ace-mcp starts on MCP Python SDK 2.x. The mcp extra now installs mcp>=1.22.0,<3. (#145) - A failed recursive child session no longer changes the parent skillbook. Each child works on an isolated copy and commits back atomically. (#143, fixes #134) …

   View on GitHub ↗
2. v0.12.0 v0.12.0 — RR/Skillbook v2 rewrite + SM hardening 2026-05-07
   Release notes excerpt

   This is the merger of two release lines that had not yet shipped to PyPI: the 0.11.0 architectural rewrite and the 0.12.0 SkillManager hardening. Skipping a separate v0.11.0 tag — v0.12.0 supersets it. 0.11.0 — Architectural rewrite - RecursiveAgent core abstraction extracted from RR (ace/core/recursive_agent.py). Generic recursive PydanticAI agent with sandbox, microcompaction, default tool set, depth-aware sub-agent registration. …

   View on GitHub ↗
3. openclaw-tracing-v0.1.1 @kayba_ai/openclaw-tracing v0.1.1 2026-04-27
   Release notes excerpt

   Initial release of the OpenClaw tracing plugin. Registers as an OpenClaw plugin (\kayba-tracing\) and ships one structured Kayba trace per agent turn — user message, full LLM input/output with thinking, tool calls, final reply — with sessionId, userId, request/response previews, and folder tagging. No client SDK changes required. Capture knobs (\plugins.entries.kayba-tracing.config\): - \captureSystemPrompt\: capture system prompt once per session (default: true) - \captureHistory\: \delta\ (default, per-session cursor — eliminates N² bloat) | \full\ | \none\ …

   View on GitHub ↗

FAQ