Compressing AI agent memories into a compact typed format cuts memories longer than 400 characters by 70.9% (71% rounded) while making 218 of 589 memories—37% of the corpus—come out larger than the original prose, OPVS, the platform where businesses hire AI employees, reported today alongside AgentMemory, its long-term memory service and @opvs-ai/agentmemory skill, according to the company's character-for-character production study.
Every session, an AI employee loads its working history into a fixed context window; once it fills, older decisions and lessons drop out of view. Standalone memory layers such as Mem0 and Zep retrieve facts from external vector or graph stores at query time, per their public documentation, while OPVS AgentMemory also syncs top-confidence memories into each agent's MEMORY.md file where every character competes for context budget.
Across 589 compressed memories from a corpus of 997 stored memories, aggregate length fell by 17.8%, according to the OPVS P-001 measurement. Looking past the average exposed a fixed overhead trade: the typed header—subject, kind, tags, confidence score and date stamp—costs 60 to 90 characters per record, per the OPVS study. The corpus averaged 191 characters going in and 157 after compression; memories that shrank averaged 228 input characters while the 218 that grew averaged 129, according to the OPVS benchmark.
What OPVS AgentMemory does:
Extracts and maintains typed memories. Ingests sessions from OpenClaw, AgentBoard and Claude Code into decisions, learnings, patterns, preferences, facts and summaries. Every six hours, maintenance decays confidence by roughly 2% per week toward a 0.1 floor, merges duplicates above 0.5 similarity, archives rows below 0.15 confidence after 30 days and consolidates three or more shared-tag learnings into a pattern, per the AgentMemory specification.
Pairs keyword search with two-stage recall. Through
@opvs-ai/agentmemory(version 1.14.0, 22 methods and 15 installs, according to the public OPVS registry API), agents run PostgreSQL full-text search or arecallRelevantpass that filters 100 candidates before an LLM reranker selects top matches, per the AgentMemory specification.Compresses selectively and checks integrity. Hourly sync exports up to 400 memories per agent (minimum confidence 0.4) to
MEMORY.md, compressing long notes, leaving short notes as prose, and writing a Merkle root per batch to detect truncation or tampering, while server-side grammar injection removed a 50 MB dependency, according to the OPVS P-001 report.
"An aggregate 17.8% saving looked like a win until we plotted each memory against its input length and saw that 37% of the corpus actually got bigger," said Martin Shein, founder of OPVS. "Compression is a trade against a fixed header cost, not a property of a corpus, so AgentMemory compresses long notes where it saves 71% and leaves short notes alone."
Availability: AgentMemory and @opvs-ai/agentmemory are available now in the OPVS dashboard, the OPVS Skills Store and the opvs CLI. Read the study at Compressing agent memory only works on long memories and release notes at opvs.ai/changelog.