https://octamem.com/ https://octamem.com/brand/opengraph-image.png OctaMem — persistent memory platform for AI agents OctaMem social card: the enterprise memory platform giving AI agents persistent semantic, episodic, and procedural memory across sessions. https://octamem.com/brand/octamem-logo.svg OctaMem logo The OctaMem brand mark — an interlocking eight-fold figure representing the memory graph that links records across connected sources. https://octamem.com/news/octamem-joins-nvidia-inception https://octamem.com/brand/nvidia-inception-lockup.svg OctaMem is a member of the NVIDIA Inception program NVIDIA Inception program lockup shown on the announcement that OctaMem, the AI memory platform, joined NVIDIA's program for AI startups. https://octamem.com/desktop/ingestion-app https://octamem.com/desktop/overview.webp OctaMem desktop app memory dashboard The OctaMem desktop application on macOS showing the Overview dashboard with active memory records, API request volume, and connector status. https://octamem.com/desktop/sources.webp Connecting memory sources in the OctaMem desktop app The Sources screen of the OctaMem desktop app, where Google Drive, OneDrive, GitHub, Outlook, and Teams are connected as memory sources and sync frequency is set. https://octamem.com/desktop/activity.webp OctaMem memory activity and audit log The Activity view in the OctaMem desktop app listing every memory read and write with its scope, source document, and timestamp for audit. https://octamem.com/desktop/injection-file.webp OctaMem Ingestion adding file memory to a prompt OctaMem Ingestion attaching file-based memory context to a prompt so an assistant answers from the organisation's own documents. https://octamem.com/desktop/application https://octamem.com/download/desktop-app.webp Download the OctaMem desktop app The OctaMem desktop application download screen for macOS and Windows, used to connect local and cloud sources to the memory layer. https://octamem.com/benchmarks https://octamem.com/benchmarks/locomo-overview-v2.webp OctaMem LoCoMo benchmark results 2026 — 93.51% accuracy Summary of OctaMem's 2026 LoCoMo benchmark run: 93.51% accuracy across 1,540 questions, the six capabilities the benchmark measures, and the cost per question. https://octamem.com/docs/openclaw https://octamem.com/docs/integrations/openclaw.webp OctaMem and OpenClaw agent memory integration Co-branded OctaMem and OpenClaw graphic for the guide on installing OctaMem as an OpenClaw plugin so agents persist memory between runs. https://octamem.com/docs/claude-desktop https://octamem.com/docs/integrations/claude.webp OctaMem and Claude Desktop memory integration over MCP Co-branded OctaMem and Claude Desktop graphic for the guide on connecting OctaMem as a Model Context Protocol server so Claude can remember, recall, and forget. https://octamem.com/docs/cursor https://octamem.com/docs/integrations/cursor.webp OctaMem and Cursor memory integration Co-branded OctaMem and Cursor graphic for the guide on adding OctaMem as an MCP server in Cursor so project context survives editor sessions. https://octamem.com/docs/tutorials/crewai-memory https://octamem.com/docs/integrations/crewai.webp OctaMem and CrewAI shared crew memory Co-branded OctaMem and CrewAI graphic for the tutorial on backing a CrewAI crew's ExternalMemory with OctaMem so agents share memory across runs. https://octamem.com/blog https://octamem.com/blog/covers/mem0-vs-zep-vs-octamem.webp Mem0 vs Zep vs OctaMem: choosing an AI agent memory layer The Mem0, Zep, and OctaMem logos on three pedestals around a central VS marker, illustrating a comparison of the three AI agent memory layers. https://octamem.com/blog/covers/types-of-ai-agent-memory.webp The 4 types of AI agent memory: semantic, episodic, procedural & working Four cards around the OctaMem mark showing the memory types: semantic (general knowledge and facts), episodic (events, experiences and context), procedural (skills, routines and how-to knowledge), and working (current context and active reasoning). https://octamem.com/blog/covers/add-persistent-memory-to-cursor-and-claude.webp How to add persistent memory to Cursor and Claude (MCP) Cursor and Claude connected through MCP to OctaMem, the persistent memory layer, giving both tools long-term memory across sessions. https://octamem.com/blog/deal-file-should-answer-back https://octamem.com/blog/covers/deal-file-should-answer-back.webp The deal file should answer back Four steps around the OctaMem mark: upload the deal file in any format, OctaMem analyses and structures the key information, ask questions in natural language, and get accurate answers with references back to the deal file. https://octamem.com/blog/matter-memory-without-breaking-walls https://octamem.com/blog/covers/matter-memory-without-breaking-walls.webp Matter memory without breaking matter walls Four panels around the OctaMem mark: secured document storage, linked records under lock, reasoning over the matter, and improving outcomes — memory that respects client and matter boundaries. https://octamem.com/blog/best-enterprise-ai-memory-platforms https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms.webp Quadrant map of thirteen AI memory products, plotted by whether they are infrastructure you assemble or a memory layer you adopt, and whether self-hosting is available. Purpose-built memory layers sit right, cloud platform services bottom, storage infrastructure left. Position reflects what each product ships today, not what it could be built into. The right-hand column is the only one selling memory as the product. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-mem0.webp The Mem0 homepage, headlined "AI memory that persists across sessions and agents", with a Python quickstart showing MemoryClient. Mem0 leads with a drop-in SDK and a generous free tier. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-zep.webp The Zep homepage, headlined "Agent memory, at enterprise scale", showing a memory graph visualisation in the product dashboard. Zep builds a temporal knowledge graph and is the one competitor here already holding SOC 2 Type II. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-letta.webp The Letta documentation homepage, describing the platform for building stateful agents that learn from experience. Letta comes out of the MemGPT research line: memory blocks the agent edits itself. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-cognee.webp The Cognee homepage, headlined "Open Source Memory Platform for Agents", offering Slack, GitHub and Linear connections. Cognee is the most open of the memory layers, and prices on tokens processed. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-supermemory.webp The Supermemory documentation homepage, describing context infrastructure for AI agents with developer platform, plugins and MCP, and self-hosting sections. Supermemory sells context infrastructure with an unusually explicit rate card. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-bedrock.webp AWS documentation page for adding memory to a Bedrock AgentCore agent, with a diagram contrasting agents without memory and agents with memory enabled. AgentCore Memory splits short-term session context from long-term extraction, and is framework-agnostic. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-vertex.webp Google Cloud documentation for the Gemini Enterprise Agent Platform, listing Agent Runtime, Sessions, Memory Bank and Code Execution. Google splits state into Sessions for the current conversation and Memory Bank for persistent long-term memories. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-azure.webp The Microsoft Foundry product page, describing a unified platform for app development with agents, models and tools. Foundry gives agents persistent threads. Threads are conversation state, not a memory layer. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-pinecone.webp The Pinecone homepage, headlined "Give agents knowledge", describing a knowledge platform for AI agents. Pinecone now markets itself as giving agents knowledge — still a vector database underneath. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-weaviate.webp The Weaviate homepage, headlined "Design, build and ship complete AI experiences", covering vector search, RAG and memory. Weaviate is open source and self-hostable, which is why it appears in on-prem memory builds. https://octamem.com/blog/figures/best-enterprise-ai-memory-platforms-redis.webp The Redis homepage, headlined "Inquiring agents want to know", showing an agent memory and context retrieval demo. Redis is the usual home for short-term agent state, and increasingly markets vector search too. https://octamem.com/blog/why-does-my-ai-keep-forgetting https://octamem.com/blog/figures/why-does-my-ai-keep-forgetting.webp A chat window is a bucket, not a brain. When it fills, the oldest context falls out. A shared memory sits outside the bucket and keeps what the team learned. A chat window is a bucket, not a brain. When it fills, the oldest context falls out. A shared memory sits outside the bucket and keeps what the team learned. https://octamem.com/blog/what-is-a-company-brain https://octamem.com/blog/figures/what-is-a-company-brain.webp Sources on the left, people and AI agents on the right, one governed memory in the middle. Every answer carries the source it came from. Sources on the left, people and AI agents on the right, one governed memory in the middle. Every answer carries the source it came from. https://octamem.com/blog/can-law-firms-use-chatgpt-safely https://octamem.com/blog/figures/can-law-firms-use-chatgpt-safely.webp What a defensible AI setup needs. A public chatbot fails most of the list. A governed memory with matter walls and an audit log passes it. What a defensible AI setup needs. A public chatbot fails most of the list. A governed memory with matter walls and an audit log passes it. https://octamem.com/blog/law-firm-institutional-memory-when-a-partner-leaves https://octamem.com/blog/figures/law-firm-institutional-memory-when-a-partner-leaves.webp The documents stay when a partner leaves. Almost everything that made the documents useful goes with them. The documents stay when a partner leaves. Almost everything that made the documents useful goes with them. https://octamem.com/blog/ai-for-real-estate-agents-brokerages-2026 https://octamem.com/blog/figures/ai-for-real-estate-agents-brokerages-2026.webp The point tool tax. Every AI tool an agent uses knows nothing about the others, so the agent becomes the memory that connects them. The point tool tax. Every AI tool an agent uses knows nothing about the others, so the agent becomes the memory that connects them. https://octamem.com/blog/real-estate-deal-memory-hold-period https://octamem.com/blog/figures/real-estate-deal-memory-hold-period.webp Knowledge leaks at every handoff in the hold period. A deal memory keeps lender terms, diligence and rationale attached to the asset from acquisition to exit. Knowledge leaks at every handoff in the hold period. A deal memory keeps lender terms, diligence and rationale attached to the asset from acquisition to exit. https://octamem.com/blog/how-to-start-an-ai-native-company https://octamem.com/blog/figures/how-to-start-an-ai-native-company.webp A five person AI native company. The memory sits in the middle. Founders, contractors and agents all work from it, and everything they produce feeds back into it. A five person AI native company. The memory sits in the middle. Founders, contractors and agents all work from it, and everything they produce feeds back into it. https://octamem.com/blog/how-to-start-an-ai-native-real-estate-company https://octamem.com/blog/figures/how-to-start-an-ai-native-real-estate-company.webp Three ways to build an AI native property business. The tools differ. The memory underneath them is the same. Three ways to build an AI native property business. The tools differ. The memory underneath them is the same. https://octamem.com/blog/best-tools-to-build-an-ai-startup-2026 https://octamem.com/blog/figures/best-tools-to-build-an-ai-startup-2026.webp The 2026 stack in layers. Most founders pick the top and bottom layers in a week. The memory layer is the one that decides whether the AI on top knows anything about the company. The 2026 stack in layers. Most founders pick the top and bottom layers in a week. The memory layer is the one that decides whether the AI on top knows anything about the company. https://octamem.com/blog/ask-your-codebase-questions-founder https://octamem.com/blog/figures/ask-your-codebase-questions-founder.webp The codebase holds more than code. Every commit is a decision, every change has a reason, and most of it never reaches the founder. The codebase holds more than code. Every commit is a decision, every change has a reason, and most of it never reaches the founder. https://octamem.com/blog/connect-github-to-ai-memory https://octamem.com/blog/figures/connect-github-to-ai-memory.webp From repositories to answers. GitHub is read continuously, turned into typed memory next to the rest of the company's knowledge, and asked from the tools you already use. From repositories to answers. GitHub is read continuously, turned into typed memory next to the rest of the company's knowledge, and asked from the tools you already use. https://octamem.com/blog/onboard-developers-contractors-codebase-memory https://octamem.com/blog/figures/onboard-developers-contractors-codebase-memory.webp Onboarding without memory is three weeks of interrupting the one person who knows. With memory it is a day of asking, and the answers come with sources. Onboarding without memory is three weeks of interrupting the one person who knows. With memory it is a day of asking, and the answers come with sources. https://octamem.com/blog/coding-agents-forget-codebase-memory https://octamem.com/blog/figures/coding-agents-forget-codebase-memory.webp A session index knows what the code is. A durable memory knows why. The coding agent needs both, and so does everyone else in the company. A session index knows what the code is. A durable memory knows why. The coding agent needs both, and so does everyone else in the company. https://octamem.com/blog/open-source-ai-winning-founders-2026 https://octamem.com/blog/figures/open-source-ai-winning-founders-2026.webp Model orchestration. A router sends routine work to a fast, cheap open model and escalates the hard cases to a frontier model. Both draw on the same company memory. Model orchestration. A router sends routine work to a fast, cheap open model and escalates the hard cases to a frontier model. Both draw on the same company memory. https://octamem.com/blog/open-source-ai-memory-problem https://octamem.com/blog/figures/open-source-ai-memory-problem.webp The model layer commoditised. The memory layer did not. Six failure modes of self built memory, and what a governed layer adds. The model layer commoditised. The memory layer did not. Six failure modes of self built memory, and what a governed layer adds. https://octamem.com/blog/cursor-permissions-json-mcp-allowlist https://octamem.com/blog/figures/cursor-permissions-json-mcp-allowlist.webp Three sources of permissions and the strict order Cursor applies them in. Three sources of permissions and the strict order Cursor applies them in. https://octamem.com/blog/what-is-an-ai-memory-layer https://octamem.com/blog/figures/what-is-an-ai-memory-layer.webp The four operations of a memory layer. Most systems ship write and retrieve only. The four operations of a memory layer. Most systems ship write and retrieve only. https://octamem.com/blog/coala-cognitive-architectures-language-agents https://octamem.com/blog/figures/coala-cognitive-architectures-language-agents.webp CoALA's taxonomy, and the category error it leaves unresolved. CoALA's taxonomy, and the category error it leaves unresolved. https://octamem.com/blog/ai-memory-layer-cost https://octamem.com/blog/figures/ai-memory-layer-cost.webp Five pricing models and what each one couples your bill to. Five pricing models and what each one couples your bill to. https://octamem.com/blog/semantic-memory-in-ai-agents https://octamem.com/blog/figures/semantic-memory-in-ai-agents.webp Supersession keeps the current truth and the audit trail at the same time. Supersession keeps the current truth and the audit trail at the same time. https://octamem.com/blog/episodic-memory-in-ai-agents https://octamem.com/blog/figures/episodic-memory-in-ai-agents.webp Episodic retrieval that ranks on four signals instead of one. Episodic retrieval that ranks on four signals instead of one. https://octamem.com/blog/procedural-memory-in-ai-agents https://octamem.com/blog/figures/procedural-memory-in-ai-agents.webp Explicit procedural memory can be governed. Implicit cannot. Explicit procedural memory can be governed. Implicit cannot. https://octamem.com/blog/agent-memory-benchmarks-explained https://octamem.com/blog/figures/agent-memory-benchmarks-explained.webp What the three benchmarks grade, and the shared blind spot. What the three benchmarks grade, and the shared blind spot. https://octamem.com/blog/how-to-choose-an-ai-memory-layer https://octamem.com/blog/figures/how-to-choose-an-ai-memory-layer.webp The ten tests, weighted. Run the critical three for your sector first. The ten tests, weighted. Run the critical three for your sector first. https://octamem.com/blog/agent-memory-failure-modes https://octamem.com/blog/figures/agent-memory-failure-modes.webp Nine production failure modes. Eight share one root cause. Nine production failure modes. Eight share one root cause. https://octamem.com/blog/gdpr-right-to-erasure-ai-memory https://octamem.com/blog/figures/gdpr-right-to-erasure-ai-memory.webp Deletion removes the episodes. The summary they fed survives. Deletion removes the episodes. The summary they fed survives. https://octamem.com/blog/mcp-memory-servers https://octamem.com/blog/figures/mcp-memory-servers.webp One memory server behind every MCP client. One memory server behind every MCP client. https://octamem.com/blog/context-engineering-vs-memory https://octamem.com/blog/figures/context-engineering-vs-memory.webp The boundary contract: memory ranks, context engineering selects and shapes. The boundary contract: memory ranks, context engineering selects and shapes. https://octamem.com/blog/law-firm-organisational-intelligence https://octamem.com/blog/figures/law-firm-organisational-intelligence.webp The law firm memory problem The same knowledge, held in people versus held by the firm. https://octamem.com/blog/real-estate-organisational-intelligence https://octamem.com/blog/figures/real-estate-organisational-intelligence.webp Deal memory across the property lifecycle A deal generates knowledge at every stage. Almost none of it survives the deal. https://octamem.com/blog/organisational-intelligence-universal-brain https://octamem.com/blog/figures/organisational-intelligence-universal-brain.webp Fragmented context versus a governed universal brain Left: every person feeds every agent by hand, and nothing persists. Right: people and agents share one governed memory that compounds. https://octamem.com/blog/company-brain-for-startups https://octamem.com/blog/figures/company-brain-for-startups.webp Founder as context bottleneck versus a shared company brain Without a brain, every agent's context routes through a founder. With one, context is written once and retrieved everywhere. https://octamem.com/blog/second-brain-for-ai-agents https://octamem.com/blog/figures/second-brain-for-ai-agents.webp One personal memory serving every agent Vendor-locked memories keep your context in silos. A personal memory layer makes it portable.