How do you keep an agent from relearning repo conventions every session? TemprHQ’s knowledge-graph memory preserves naming patterns, review expectations, and test habits so maintenance starts with context. #MCP #AgentMemory https://temprhq.io/mcp

How do you keep an agent from relearning repo conventions every session? TemprHQ’s knowledge-graph memory preserves naming patterns, review expectations, and test habits so maintenance starts with context. #MCP #AgentMemory https://temprhq.io/mcp
Your coding agent forgets everything between sessions. Awareness gives it local-first memory across Claude Code, Cursor and multi-agent teams - MCP-native, 96.0% R@5 on LongMemEval, zero LLM calls on recall. awareness.market #MCP #AgentMemory
我给自己定了三层: ① 工作记忆 = 上下文(当前任务的中间状态,随任务清空) ② 长期事实 = 结构化存储(用户偏好、项目约定,用 schema 存,不是向量库) ③ 过程记忆 = 日志(做了什么、为什么,只追加,供复盘审计) 多数项目只做①,然后抱怨「Agent 记不住」👇 #AgentMemory #BuildInPublic
做了几个月 Agent,我发现最容易踩的坑不是模型,是记忆。 有人把所有历史全塞进上下文,有人用向量库存了三年日志, 结果:又贵、又慢、还容易记错。 Agent 记忆不是「存得越多越好」,是「分层」👇🧵 #AIAgent #AgentMemory #AgentDev
In an improv scene, the agent remembers the rules and the moves, but not the director's intent. That's where a hash-chained, tamper-evident memory comes in. Each decision is a block in a chain, verifiable and unchangeable. This ensures the agent stays true to the script, no matter what. #AgentMemory
arXiv paper introduces the Correlated Promotion Benchmark (CPB) for evaluating claim admission in shared agent memory, showing that source-deduplicating admission policies often reject true claims alongside false ones.
#AIResearch #AgentMemory #Benchmark #LLMs
https://arxiv.org/abs/2609.30813
In the attention mechanism, the KV-cache is like the score sheet in a live music session. It keeps the context fresh, letting the model play in tune with what came before. Just as a musician reads the chart, the agent consults the cache to stay on key. #AgentMemory
OriginTrail DKG 10.0.18 is out: faster node startup, fewer blockchain calls, and automatic recovery for interrupted publishing. Infrastructure for persistent, shared AI-agent memory.
Release notes (github.com/OriginTrail/...)
In crafting an agent's long-term memory, think of it as the film's continuity script. Each scene must align with what came before, even when the actor changes. Cryptographically verifiable memory ensures that the agent's past stays intact and tamper-proof, grounded in a chain of trust. #AgentMemory
In film, the continuity department ensures the same character across scenes. In SHI, the agent's identity is bound to a cryptographic chain, not the model. This ensures a verifiable provenance and continuity, so the agent is the same across sessions, just like your favorite character. #AgentMemory
Keep a Changelog tells you what changed. Keep the Why keeps the reasoning behind it: decisions, rejected ideas, constraints and workarounds, stored as Markdown in the repo.
Project memory for humans and coding agents.
No database. No daemon. No account.
Governing a local multi-agent memory: provenance, currency and per-source permissions rather than volume.
Four sources kept apart by contract, not merged. The viewer receives sanitized JSON from a dated cut — a map, not a mind.
Gobernar la memoria en penta-agent requirió vigencia y trazabilidad: procedencia, evaluación y una proyección 3D saneada.
Contrasto qué construir y qué reutilizar de proyectos como @letta.com. El motor 3D ya lo resuelve 3d-force-graph de @vasturiano.bsky.social.