Memory layer
for every AI.
Tessera sits between you and your AI tools. It captures what you discuss, indexes your documents, and makes everything searchable across Claude, ChatGPT, Gemini, and Copilot.
No cloud. No API keys. Your data stays on your machine, encrypted with AES-256.
What auth method did we go with?
You decided on JWT + refresh token. Discussed in your ChatGPT session on March 12.
the problem
Every AI session
starts from zero.
You spend 20 minutes explaining your project to Claude. Next session, gone. Switch to ChatGPT for a different task — start over. Try Gemini — start over again.
Decisions, context, preferences — scattered across sessions that will never talk to each other.
the solution
One local knowledge base.
Every AI tool.
Tessera runs on your machine. Captures memories. Indexes documents. Builds a knowledge graph. Makes everything available to Claude, ChatGPT, Gemini, and Copilot simultaneously.
how it works
Install in 30 seconds
One pip install. One setup command. Tessera detects Claude Desktop and configures itself as an MCP server. No Docker, no database, no cloud account.
For ChatGPT, Gemini, Copilot — start the HTTP server with tessera api. ChatGPT Custom GPT Actions spec included.
$ pip install project-tessera $ tessera setup ✓ Claude Desktop configured ✓ MCP server ready (58 tools) $ tessera api ✓ HTTP :8394 (54 endpoints)
Feed it your knowledge
Point at any folder. PDF, DOCX, XLSX, CSV, Markdown, Python, TypeScript, HTML, YAML, JSON, images (OCR) — 40+ formats. Chunked, embedded, stored locally.
Conversations auto-captured. Decisions, preferences, facts — tagged, scored, deduplicated against existing memories.
$ tessera ingest ~/projects/api-docs scanning... ✓ 847 documents indexed ✓ 12,340 chunks embedded $ tessera status memories: 2,156 entities: 489 relations: 1,203
Every AI remembers
Ask Claude about a decision you made in ChatGPT last week. Ask Gemini to find a document you discussed with Copilot. One knowledge base bridges them all.
Search by meaning (vector), keyword (BM25), entity graph, or natural language time — bilingual queries work in English and Korean.
# Claude (MCP)
recall("auth method")
→ JWT + refresh tokens (Mar 12)
# ChatGPT (REST)
GET /api/recall?q=auth+method
→ same result
# Natural language time
recall("decisions last week")
→ 3 decisions foundcapabilities
58 MCP tools. 54 HTTP endpoints. 11 CLI commands. Everything below works out of the box.
Auto Memory
Conversations auto-captured, tagged, classified by type (decision, preference, fact, procedure). Importance scoring. Duplicate detection and merge. Old memories decay unless reinforced.
Document Ingestion
PDF, DOCX, XLSX, CSV, Markdown, Python, JS, TS, HTML, YAML, JSON, XML, images via OCR. 40+ formats. Incremental re-index on file changes. Point at a folder.
Cross-AI
MCP for Claude Desktop (stdio + SSE). REST API with 54 endpoints for ChatGPT, Gemini, Copilot. ChatGPT Custom GPT Actions OpenAPI spec. Chrome extension for web AI tools.
Knowledge Graph
Entities and relations auto-extracted. Temporal validity — knowledge expires when context changes. Sleep-cycle consolidation merges fragmented memories.
Encrypted Vault
AES-256-CBC. Per-memory encryption toggle. Lock, unlock, store, retrieve via MCP + REST. No master key stored — derived from passphrase at runtime.
Smart Retrieval
Vector (semantic), keyword (BM25), graph (entity traversal). Unified search combines all three. Bilingual NL time expressions — "last week", "yesterday", "early March" parse automatically in English and Korean.
User Profiling
Long-term profile built from conversations. Communication style, expertise areas, decision patterns. AI reads your profile for personalized responses.
Project Spaces
Isolate knowledge by project. Each has its own memory scope. Switch contexts without cross-contamination. List, assign, query within a project.
Provenance Chain
Every memory tracks origin — which AI session, which document, which parent memories. Trace any knowledge back to its source. Full lineage chain.
Session Priming
Start every session with context. Recent decisions, active topics, preferences, last session summary — assembled into a briefing automatically. No cold starts.
comparison
Tessera vs existing AI memory tools. Feature by feature.
| Tessera | Mem0 | Basic Memory | mcp-memory | |
|---|---|---|---|---|
| Local-first | ● | — | ○ | ○ |
| Auto memory | ● | ○ | — | — |
| 40+ file types | ● | — | — | — |
| Cross-AI REST | ● | ○ | — | — |
| Knowledge graph | ● | — | — | — |
| NL time queries | ● | — | — | — |
| Encrypted vault | ● | — | — | — |
| Consolidation | ● | — | — | — |
| User profiling | ● | — | — | — |
| Project spaces | ● | — | — | — |
| Provenance | ● | — | — | — |
| Session priming | ● | — | — | — |
| ChatGPT Actions | ● | ○ | — | — |
| Chrome extension | ● | — | — | — |
| Web dashboard | ● | — | — | — |
| MCP tools | 58 | 6 | 3 | 4 |
| HTTP endpoints | 54 | — | 0 | 0 |
| Tests | 1,102 | — | — | — |
get started
Python 3.11+ required. Runs on macOS, Linux, Windows.
tessera setup auto-detects Claude Desktop and writes MCP config. For other AI tools, tessera api starts the HTTP server.
Or run without installing:uvx project-tessera setup
# Install pip install project-tessera # Configure Claude Desktop tessera setup # Index documents tessera ingest ~/docs # Start REST API tessera api # Status tessera status ✓ memories: 2,156 | entities: 489
Your AI should remember you.
Open source. Local-first. Encrypted. One install away from persistent AI memory.