# past.dev blog

> Engineering, research and company notes from past.dev. Every post is also available as Markdown by appending .md to its URL.

- [Why RAG gets less accurate as your data grows](https://past.dev/blog/why-rag-gets-less-accurate) (2026-10-09, Research): On BEAM at 10M tokens, the RAG in the BEAM paper scores 21.1% to 24.9% and past.dev 85.0%. Why RAG accuracy drops as your data grows.
- [Introducing past.dev](https://past.dev/blog/introducing-past-dev) (2026-09-30, Announcements): past.dev, a memory layer for AI agents, is available today. An agent asks it for context and gets current facts with dated sources, limited to what it may see.
- [You'll never reach 85% on BEAM10M using .md files](https://past.dev/blog/md-files-vs-memory-api-beam-10m) (2026-09-30, Engineering): Why agent memory kept in .md files falls short on BEAM at 10M tokens, and how the past.dev Memory API handles changing facts, late records and permissions.
- [How to Add Long-Term Memory to ChatGPT](https://past.dev/blog/add-long-term-memory-to-chatgpt) (2026-09-03, Research): Every documented way to add long term memory to ChatGPT: native memory and its limits, projects, custom GPTs, MCP apps in developer mode, and the API.
- [How to Add Long-Term Memory to Claude](https://past.dev/blog/add-long-term-memory-to-claude) (2026-09-03, Research): Every documented way to add long term memory to Claude: built-in memory settings, Projects, custom connectors via remote MCP, and memory MCP servers.
- [How Codex's memory works](https://past.dev/blog/codex-memory) (2026-09-03, Research): Codex memory, from OpenAI's docs: AGENTS.md loading order, the generated memories directory, cloud container caching, CLI sessions, and controls.
- [Memory in DeepSeek and Windsurf](https://past.dev/blog/deepseek-windsurf-memory) (2026-09-03, Research): DeepSeek memory and Windsurf memory from official docs: what persists across sessions, which controls exist, and what changed in 2026.
- [How to Delete ChatGPT Memory and Manage What It Keeps](https://past.dev/blog/delete-chatgpt-memory) (2026-09-03, Research): How to delete ChatGPT memory: remove single saved memories, clear everything, what deleting a chat leaves behind, and the settings that turn memory off.
- [Does Claude Have Memory? Settings, Controls, Limits](https://past.dev/blog/does-claude-have-memory) (2026-09-03, Research): Yes, Claude does have memory. The exact settings to view, edit, pause, reset, or turn off Claude memory, plus incognito chats and Team admin controls.
- [Does Gemini Have Memory? Settings and Controls](https://past.dev/blog/does-gemini-have-memory) (2026-09-03, Research): Yes, Gemini does have memory. Where the Memory toggle lives, how deletion works through Gemini Apps Activity, Temporary Chats, and availability rules.
- [How Grok's memory works](https://past.dev/blog/grok-memory) (2026-09-03, Research): Grok memory explained from official sources: what Grok remembers across conversations, the controls xAI documents, and how deletion works.
- [How Hermes Agent's memory works](https://past.dev/blog/hermes-agent-memory) (2026-09-03, Research): Hermes agent memory, from the official docs: two bounded files with hard caps, capacity errors, SQLite session search, and external providers.
- [Mem0 vs Cognee: Extracted Facts or a Knowledge Graph](https://past.dev/blog/mem0-vs-cognee) (2026-09-03, Research): Mem0 vs Cognee compared: extracted per-user facts against knowledge graph construction. Search modes, pricing, deployment, and Cognee vs Graphiti.
- [Mem0 vs LangMem: Library or Service for Agent Memory](https://past.dev/blog/mem0-vs-langmem) (2026-09-03, Research): Mem0 vs LangMem compared on memory model, temporal handling, retrieval, deployment, and pricing, with a decision path for LangGraph teams.
- [Mem0 vs Letta: Memory Layer or Stateful Agents](https://past.dev/blog/mem0-vs-letta) (2026-09-03, Research): Mem0 vs Letta compared: an extraction memory layer against hosted stateful agents with memory blocks. Pricing, deployment, and the MemGPT rename.
- [Mem0 vs Supermemory: Two Memory APIs Compared](https://past.dev/blog/mem0-vs-supermemory) (2026-09-03, Research): Mem0 vs Supermemory compared: extracted user memories against document memory with connectors, what search returns, pricing, and self-host options.
- [Mem0 vs Zep: Extraction or a Temporal Graph](https://past.dev/blog/mem0-vs-zep) (2026-09-03, Research): Mem0 vs Zep in practice: extracted fact strings against a temporal knowledge graph, what each returns, pricing meters, and the Zep vs Graphiti split.
- [Mem0 vs Zep vs Letta: Building the Memory Shortlist](https://past.dev/blog/mem0-vs-zep-vs-letta) (2026-09-03, Research): Mem0 vs Zep vs Letta for the agent memory shortlist: extraction memory, a temporal knowledge graph, or hosted stateful agents, plus the wider field.
- [How OpenClaw's memory works](https://past.dev/blog/openclaw-memory) (2026-09-03, Research): OpenClaw memory, from the official docs: the Markdown files that persist, what loads at session start, the documented caps, and the dreaming sweep.
- [How Perplexity's memory works](https://past.dev/blog/perplexity-memory) (2026-09-03, Research): Perplexity memory explained: what Memory and Brain store, when answers draw on them, and how to view, edit, or delete memories in Settings.
- [How ChatGPT Memory Works](https://past.dev/blog/chatgpt-memory) (2026-09-02, Research): How ChatGPT memory works as of September 2026: saved memories, reference chat history, the memory summary, what memory full means, and how to clear it.
- [Claude Code Memory: How CLAUDE.md Actually Works](https://past.dev/blog/claude-code-memory) (2026-09-02, Research): How Claude Code memory works: the CLAUDE.md hierarchy, auto memory files and caps, session resumption, and where file-based facts go stale.
- [Claude Memory: How It Works and What It Stores](https://past.dev/blog/claude-memory) (2026-09-02, Research): How Claude memory works in September 2026: what gets stored, project scoping, plan availability, user controls, and the limits Anthropic documents.
- [How Copilot Memory Works: Microsoft and GitHub](https://past.dev/blog/copilot-memory) (2026-09-02, Research): Copilot memory means two different systems. A teardown of Microsoft Copilot and GitHub Copilot memory: storage, retention, controls, deletion.
- [How Cursor Memory Works: Rules and Memories](https://past.dev/blog/cursor-memory) (2026-09-02, Research): Cursor memory from the official docs: rules vs the memories feature, what persists across chats and machines, plus controls, deletion, limits.
- [How Gemini Memory and Personalization Work](https://past.dev/blog/gemini-memory) (2026-09-02, Research): How Gemini memory works as of September 2026: past chat recall, Instructions for Gemini (saved info), Personal Intelligence, and deletion controls.
- [What a good AI memory benchmark measures](https://past.dev/blog/ai-memory-benchmark-guide) (2026-09-01, Research): An AI memory benchmark must test changing facts, ordering, and abstention on complete splits. The ten abilities, and what to demand from any score.
- [Cutting LLM token costs with memory](https://past.dev/blog/llm-token-costs) (2026-09-01, Research): LLM token costs scale with how much history you send per question. The arithmetic of context stuffing, what caching changes, and what memory changes.
- [How memory accuracy degrades with scale](https://past.dev/blog/memory-accuracy-at-scale) (2026-09-01, Research): Memory accuracy measured at 100K, 500K, 1M and 10M tokens: the published splits, which abilities hold at scale, and where systems actually lose accuracy.
- [What is a company brain? Timeline and requirements](https://past.dev/blog/company-brain) (2026-08-30, Research): A definition of the company brain, its main components, and the memory functions required to keep company information current.
- [Giving AI agents a past](https://past.dev/blog/giving-ai-a-past) (2026-08-30, Research): AI agents start every run without a past. What giving AI a past means: event time, sources, supersession, and how agents track the past accurately.
- [What is an LLM wiki and where does it break?](https://past.dev/blog/llm-wiki) (2026-08-30, Research): A definition of the LLM wiki, a standard implementation using Markdown and git, and its limitations for facts that change.
- [Mem0 alternatives for agent memory in 2026](https://past.dev/blog/mem0-alternatives) (2026-08-30, Engineering): A comparison of mem0 alternatives including Zep, Letta, Supermemory and past.dev by architecture, deployment, inputs and changed-fact handling.
- [What agentic memory is and how to build it](https://past.dev/blog/agentic-memory) (2026-08-28, Research): Agentic memory explained: the four memory types, why agents need storage outside the context window, and how a temporal knowledge graph works.
- [How AI agents read a website](https://past.dev/blog/website-agents-can-read) (2026-08-26, Engineering): How AI agents read a website through HTML, Markdown and llms.txt interfaces, with tests for every published format.
- [How we benchmark memory](https://past.dev/blog/measuring-memory-honestly) (2026-08-25, Research): How past.dev benchmarks memory with corrected answer keys, strong baselines, multiple judges, and cost reported with accuracy.