One Memory, Everywhere
Gnosis gives every AI assistant the same persistent memory. 14+ clients: Claude, ChatGPT, Cursor, VS Code, Copilot CLI, and anything that speaks MCP. Switch tools, switch providers, switch devices. The memory follows.
Each memory is typed (fact, decision, task, preference, path) with searchable topics and structured metadata. Not flat text. Preferences load on session start. Tasks carry checkboxes that persist across conversations. Store a memory in Claude Code on your laptop. Open ChatGPT on your phone. It's there.
Data portability
- One-click JSON export of your entire memory corpus, anytime
- No proprietary formats. Structured JSON with standard vector embeddings
- Import from other services planned for Claude.ai, ChatGPT, and plain text
Search by Meaning, Not Keywords
At session start, Gnosis loads a topic landscape: a compressed snapshot of everything you've stored. Active tasks, behavioral preferences, and your most important context, loaded in one call. Nothing random.
When specifics are needed, semantic search finds memories by meaning. "Database performance" surfaces query optimization notes that never use the word "performance." Results in under 100ms. 100+ languages, no configuration.
The AI gets compressed previews first, then retrieves full content only for the 2-3 results it actually needs. In most memory systems, the system guesses what's relevant. In Gnosis, the AI decides.
Consolidation: executive summaries from dozens of memories
When a topic accumulates enough memories, the AI can distill them into an executive summary. The summary surfaces near the top of future searches. New sessions get instant context without reading 30 individual memories.
Memory Quality
Quality Without a Server-Side AI
No server-side LLM. No content filter. No invisible quality review. The MCP tool descriptions are detailed enough that AI clients write structured, specific, searchable memories on their own. Even 8B-parameter models get it right. 99.8% of 2,095 audited memories rated B+ or better.
Gnosis tool descriptions encode creation guidelines, topic conventions, and quality heuristics directly into the MCP schema. Good structure is the path of least resistance.
Two-tier deduplication
- Hash fast-path catches exact text matches instantly. Zero overhead
- Semantic similarity compares new memories against existing ones. Near-duplicates are flagged and returned to the AI, which can review and edit the existing memory, or force-store if the new one is genuinely distinct
Nothing is silently dropped. The AI makes the call, and you see it happen in your conversation.
Private by Architecture
When you evaluate a memory service, ask one question: can the operator read my data? If the answer is "we won't" instead of "we can't," that's a policy. Policies change.
"Can't, Not Won't"
Gnosis encrypts memories at rest using AES-256 with per-user keys derived from your credentials. Keys exist only in memory during active sessions. Never on disk. Never in a database. Never accessible to an admin panel. There is no master key. We cannot decrypt your memories. Cryptography, not policy.
Gnosis stores exactly what the AI tells it to store. No server-side LLM decides what's "important enough" to keep. No invisible filtering. No editorial decisions.
Compliance
- GDPR Article 34(3)(a): encrypted data breaches do not require individual notification
- US state safe harbors: multiple breach notification laws exempt encrypted data
- Privacy by design: GDPR Article 25 compliance is architectural, not bolted on
Designed for SOC 2 and HIPAA from day one.
Shared Knowledge, Not Shared Files
Planning a trip with your spouse? Building something with a friend? Working on code with a teammate? Everyone uses a different AI, and none of them talk to each other.
One Collection, Every Member's AI
Create a shared collection and invite anyone by email. Every AI tool each person uses can read and write to the same knowledge. No sync, no file sharing, no merge conflicts. Two people on different providers share knowledge in under 30 seconds.
Two collection types: collaborative (everyone reads and writes) and knowledge packs (one person curates, others read). Search automatically includes every collection you belong to. No extra configuration needed.
Agent Identities
Different AI workflows need different memory spaces. A code-review agent shouldn't dump findings into your personal memories. A research bot shouldn't see your debugging notes.
Isolated Memory Per Persona
Each agent identity gets its own memory namespace under your account. You control visibility: grant an agent read access to your personal memories, or keep it fully isolated. Agents join shared collections for team coordination.
Three-ring access model
- Agent namespace — the agent's own working memory, isolated by default
- Owner memories — controlled by
owner_access: none, reader, or member - Shared collections — read/write per collection role
Search queries all accessible layers automatically. No per-query configuration needed.
How agents coordinate → · Plus plan → · agent_manage reference →
Tasks That Survive Between Sessions
Gnosis tasks are memories with checkboxes. Searchable, typed, persistent. Start a task Tuesday, come back Wednesday. The checkboxes, status, and linked findings are exactly where you left them.
Structured Work Tracking
Status workflow: pending, active, blocked, review, done. In shared collections, task state changes are visible to every member automatically. One person's agent creates a task. Another person's agent picks it up. No handoff meeting. No Slack thread.
Task surgery — precision edits without rewriting
- toggle — flip checkboxes by index
- set_status — advance the workflow
- output capture — attach a finding that links back to the task. Search for the task ID, find the task and all linked findings together
Signals
Send a memory or task to another person or agent so they have context for what you need. Send your coworker a decision. Send your spouse a task. Send an agent the next piece of work with the right checkbox already queued up for its role. Recipients see incoming signals on their next session start.
Trigger Agents Across Machines
Signals enable distributed agent frameworks that would normally require a daemon on the same machine. An agent framework can poll for signals at near-zero cost to both the user's infrastructure and ours. When a signal arrives, the agent wakes, reads the referenced memories, and acts. This is how you build coordination between agents that never run at the same time or on the same hardware.
Secure by design
Signals contain only memory IDs. No content, no metadata, no relationship information. The recipient must already have access to the shared collection to read the referenced memories. Even in a complete breach, signals cannot be used to reverse-engineer relationships, collection memberships, or memory content. Nothing is transmitted except opaque IDs. Signals auto-expire after 48 hours.