Ideas, tutorials, and technical deep-dives on AI memory.
A memory system for AI looks like one thing, but it holds three: facts, identity and procedure. Which one the AI may edit on its own, why skills change the rules, and how they enter through the door.
Read more → ReflectionA persistent, local memory that plugs into any AI changed how I work day to day: onboarding any model in seconds, multi-AI deliberation, passive capture and time-aware recall. Eight concrete use cases.
Read more → ReflectionI built a shared table where Claude, ChatGPT and DeepSeek deliberate in real time over the same local memory. The first day caught two bugs, and I haven't seen this running anywhere else.
Read more → ReflectionClaude, ChatGPT and DeepSeek now share and independently audit the same local memory. I shipped the changes all three agreed on, and EIDARA moves from building to measuring.
Read more → ReflectionThe third MCP client on my local memory is from a different AI vendor: OpenAI's Codex, alongside Anthropic's Claude. It worked, which is evidence for EIDARA's vendor-agnostic design.
Read more → ReflectionI asked a new Claude model, Fable, to audit EIDARA end to end. Seven structural upgrades came out of it, from confidentiality zones to a defense against silent drift.
Read more → ReflectionTwo weeks after the senses went live, the work turned to replacing each rough edge with a structural fix that won't break the same way twice. An update on where EIDARA is.
Read more → ReflectionI gave my AI memory three senses: voice, typing and screen text. The most important feature is the off switch, plus what two weeks of use showed.
Read more → TechnicalI built an AI memory that reads my screen without looking at it. Why I rejected computer vision for the accessibility tree.
Read more → TechnicalTwo AIs disagreed about a fact in my memory. I built four ways for that conflict to resolve, and only one of them involves me.
Read more → ReflectionAn outside agent audited my memory system and found five debts the dashboard wasn’t showing. The dashboard was lying.
Read more → ReflectionThe auditor reported the same issues daily that the curator was already fixing. The principle: every failure becomes a prevention rule, not a patch.
Read more → TechnicalOne memory neuron silently accumulated 156 facts that didn’t belong. A cascade misrouting bug, and how a single guard fixed it overnight.
Read more → TechnicalReal data from two weeks of autonomous AI memory: what worked, what broke and what I changed, with the numbers behind each decision.
Read more → TechnicalMultiple AIs contribute atomic facts through MCP, and one local curator decides what gets integrated. DARA stopped being a folder and became a service.
Read more → ReflectionI activated an autonomous curator. Five hours later, five false positives. Two weeks later, zero errors. Here's what happened.
Read more → Reflection140 cycles, zero errors. A sleep-time agent curating local-first AI memory: what happened, what broke, and what comes next.
Read more → ReflectionWe activated the autonomous side of DARA. Five hours later, five false positives. Here's what we did about it.
Read more → Reflectionv2.0 makes the system self-correcting. Plus: atomic memory and desktop context.
Read more → ReflectionWhy I abandoned hierarchical folders for a compiled, flat approach to AI memory.
Read more → ReflectionOne AI with full context outperforms six agents without it.
Read more → ReflectionMECE is a thinking tool, not a memory tool. Here's why I had to unlearn it.
Read more → ReflectionWhy access permissions create bottlenecks, and what works instead.
Read more → TechnicalCompile your context instead of summarizing it.
Read more → TechnicalNo manual cleanup. No curator. The system maintains itself through consensus.
Read more → Technical52 tests each, no human intervention. Here's what happened.
Read more → TutorialNo technical skills required. Download, hand it to any AI, done.
Read more → TutorialAn agent is a .md file. Share it like any document.
Read more →