IBM ALTK-Evolve Cuts Agent Context Costs by Up to 85% Over ACE
IBM Research introduced ALTK-Evolve, an open-source agentic memory library that replaces full-playbook prompt injection with selective guideline retrieval. On the AppWorld benchmark, it matches or exceeds ACE accuracy at 40% to 14% of the inference token cost.
Impact: High
Why it matters
Developers building ReAct or multi-step AI agents can dramatically lower API bills by retrieving curated guidelines instead of injecting raw execution histories into every step.
TL;DR
- 01Avoid injecting monolithic instruction playbooks into every prompt step for multi-step agents.
- 02Store distinct operational guidelines with support counts to track rule confidence without degrading detail.
- 03Use task-specific guideline retrieval to cut agent context costs by 60% to 85%.
Key facts
- Inference Cost (Strong Model)
- ~40% of ACE
- Inference Cost (Weak Model)
- ~14% of ACE (1/7th)
- Benchmark Split
- AppWorld test_normal (168 tasks)
- Evaluated Base Models
- DeepSeek-V3.2, gpt-oss-120b
Context Delivery vs Playbook Injection
Both ACE and ALTK-Evolve agree that hard-won agent lessons should not be compressed into brief summaries, as brevity bias and context collapse erase critical operational nuance. However, ACE sends its full playbook on every reasoning step. ALTK-Evolve treats context delivery as a dynamic dial, serving a small fixed core of high-support guidelines plus task-specific rules chosen via cosine or LLM-guided selection.
AppWorld Benchmark Performance
On the test_normal split of AppWorld using DeepSeek-V3.2, ALTK-Evolve achieved higher Task Goal Completion across Easy, Hard, and Overall difficulty tiers at approximately 40% of ACE's inference cost. On gpt-oss-120b, ALTK-Evolve matched ACE's accuracy (56.0 vs 54.8 TGC) while reducing inference costs to approximately 1/7th (~14%).
Try it in 2 minutes
from altk_evolve import MemoryStore, Retriever
store = MemoryStore.from_trajectories(trajectories)
retriever = Retriever(store, max_guidelines=5)
guidelines = retriever.get_guidelines(task=task_description)python
✓ When to use
- Multi-step ReAct coding or infrastructure agents that frequently interact with external APIs.
- Agent fleets facing strict context token budgets or high API costs.
✕ When NOT to use
- Single-turn chat assistants or tasks without multi-step tooling dependencies.
- Stateless workflows where past execution history offers zero transferable logic.
What to do today
- Audit existing agent prompt templates to separate static system instructions from retrieved task rules.
- Test the ALTK-Evolve Python library for extracting and retrieving guidelines from agent execution logs.
Sources