L1Working memory
System files
The rules your project always honours — read first, every session.
Markdown like MEMORY.md and CLAUDE.md. Human-readable, git-friendly, zero overhead.
View benchmarks
The memory layer
The right memory in every prompt — so your AI answers from context, not tool calls.
See how it worksThe differentiator
Same memory, every source
Other memory layers index chat history. TES indexes code, Slack, Gmail, calendar, and docs — and links them. Read a function, surface the Slack thread that decided it.
Six layers, one engine
Six layers, each tuned for a different kind of recall. L2 (HybridRAG) orchestrates them into one ranked result.
L2Reasoning
HybridRAG
The orchestrator. Asks the other layers, then fuses their answers into one ranked list.
Graph + vector + full-text, combined via reciprocal rank fusion and confidence scoring.
L3Semantic memory
Knowledge graph
Who decided what, with whom — recall by relationship, not just resemblance.
Hyperedges link people, decisions, and code in a single multi-hop traversal.
L4Episodic memory
Vector store
Semantic recall — different words, same meaning.
4,096-dimensional embeddings, chunked and reranked by a cross-encoder.
L5Social memory
Comms layer
The conversation behind the code — the thread, email, or meeting that decided it.
Slack, Gmail, and calendar, linked back into the knowledge graph automatically.
L6Procedural memory
Document store
Runbooks, RFCs, and PDFs as first-class context — not file dumps.
Full-text search with cross-encoder reranking. Rarely-read docs decay over time.
Cognitive view
Hot memories consolidate.
Cold ones decay.
Six layers say where memory lives. Four cognitive types say how it's used — each with its own capacity and decay rate.
Episodic
Recent events and turns
capacity 10,000
decay 0.05
Semantic
Consolidated knowledge
capacity 5,000
decay 0.001
Procedural
How-to and runbooks
capacity 2,000
decay 0.0001
Working
Temporary scratchpad
capacity 500
decay 0.2
Retrieval precision
Atomic facts,
not paragraphs.
Raw turns get distilled into atoms — standalone facts, back-pointed to source. Search ranks atoms above verbose turns.
Toggle distillation to see the difference
tes-memory — retrieval
why did we drop redis for sessions?
We leaned on Redis for a bunch of things last year — the queue worker rollout in March, a long stretch tuning eviction policies and TTLs, the Memcached vs Redis benchmark for the rate-limiter, and a short-lived Redis Streams pilot for the event bus that got rolled back. Sessions came up in a few of those threads too.
Start using it
One env var or one plugin install
Same memory layer for proxy and plugin. Free tier: 1M proxied input tokens / month.