TES Module · Deep Memory · Updated March 29, 2026
Your agent forgets everything. We fixed that.
A 7-layer memory architecture inspired by how biological brains actually work — from hippocampal replay to cortical consolidation.
Every memory operation is an immutable event on the TES spine. Store, recall, decay, consolidate — all auditable, all replayable.
Memory layers
7
Embedding dimensions
4096
Recall vs vector-only
+47%
Graph traversal
<30ms
The Problem
Every conversation starts from zero
LLMs have no memory. Every session is a blank slate. Context windows are finite, expensive, and vanish when the session ends. Your agent can't learn, can't grow, can't remember what happened yesterday.
"The bottleneck isn't intelligence. It's continuity."
"Memories are first rapidly encoded in the hippocampus, then gradually consolidated into neocortical networks over time."
— McClelland, McNaughton & O'Reilly, "Why There Are Complementary Learning Systems in the Hippocampus and Neocortex," Psychological Review, 1995. The neuroscience foundation behind Agent Memory's 7-layer architecture.
The Stack
Seven layers. One unified memory.
Each layer handles a different aspect of memory. Information flows down, context flows up. Inspired by complementary learning systems in neuroscience — fast acquisition in the hippocampus, slow consolidation in the neocortex. According to research published in Psychological Review, this dual-system architecture is why biological memory achieves both rapid learning and long-term stability.
L2 · HybridRAG Orchestrator
The brain of the memory stack. Fuses results from the knowledge graph, vector search, and system files using confidence scoring and reciprocal rank fusion. Graph context informs vector search — sequential, not parallel.
const results = await tes.searchMemories({
query: "What did we decide about auth?",
layers: ["semantic", "episodic"],
min_score: 0.7,
});
// → ranked results with layer + confidenceGet Started
Enable deep-memory as a TES module
Deep Memory is a built-in TES module. Enable it from your dashboard or via the API — no separate installation required. It extends the TES event log with persistent, searchable, decaying memory.
Community
Free
Apache 2.0 · L0–L4
- ✓ Platform adapters (Claude, GPT, local)
- ✓ System files & session state
- ✓ HybridRAG orchestration
- ✓ Knowledge graph
- ✓ Vector search (4096-dim)
- ✓ 100% local, 100% private
Pro
Contact us
L5–L6 + Priority Support
- ✓ Everything in Community
- ✓ Communications layer (email, chat)
- ✓ Document store with reranking
- ✓ Smart orchestration & prioritisation
- ✓ Multi-agent memory sync
- ✓ Priority support & onboarding
Open-Source SDK
Connect in two lines of code
The open-source SDK wraps your LLM client and automatically routes memory through the local 4-layer stack. Upgrade to the hosted TES platform to unlock 7 layers including communications ingestion and document store.
import { TESClient } from "@pentatonic/ai-agent-sdk";
import Anthropic from "@anthropic-ai/sdk";
const tes = new TESClient({ apiKey: process.env.TES_API_KEY });
const ai = tes.wrap(new Anthropic());
// Every call now stores episodic memory automatically
const response = await ai.messages.create({
model: "claude-opus-4-6",
messages: [{ role: "user", content: "What did we decide last week?" }],
});Frequently asked questions
Agent Memory
Give your agents the memory they deserve
Start with the open source core. Upgrade to Pro when you need communications and document ingestion.