LangChain Just SOLVED The AI Memory Problem..

LangChain shipped Open Wiki, Weaviate shipped Engram, and Skill Composer jumped 33 points on Skills Bench — the agent memory layer just got engineered properly.

## Open Wiki: a Confluence for agents LangChain open-sourced Open Wiki, a CLI tool proposed by Sydney Runkle that generates and maintains agent-consumable documentation for a codebase. You install it via npm and run it inside a repo; your coding agent then references the wiki instead of scanning every file blind. The framing is concrete: stop handing a new hire a 500-page Slack archive and start handing them a memo — agents deserve the same.

## Engram: memory as a reconciling database Weaviate's Engram replaces the typical "dump notes into a folder" pattern (the failure mode Boxmining has hit on Open Claw and Hermes). It extracts candidate memories, transforms them against what already exists, resolves contradictions, then commits the clean result. The canonical example: if you were an engineer last week and a CEO today, Engram rewrites the stale record and keeps the history instead of storing both. Pair it directly with Open Wiki for wiki + reconciled long-term store.

## Skill Composer: 33 points on Skills Bench Instead of dumping every tool into a prompt and hoping the model picks one, Skill Composer treats skill selection as a joint auto-regressive composition problem. Result: over 33 percentage points improvement on the Skills Bench Benchmark versus no-skill-selection baselines. Deep Agents now also support recursive LM workflows where a parent dispatches sub-agents — Harrison Chase connected these to agentic map reduce patterns for cost-effective security-audit sweeps.

## What changes for you The shift is from retrieval-only memory to reconciliation + maintenance, from "give the model everything" to structured workflows. Expect permission-aware, governed memory in enterprise settings rather than "a folder of markdown files sitting on a laptop somewhere."