Token-Efficient. Context-Aware. LLM-Selective.
Transform raw conversational data into structured, optimized intelligence. Dramatically reduce token usage, improve RAG precision, and ensure LLMs are only invoked when truly necessary.
Observe the deterministic L3-Condenser in action. No LLM magic—just rigorous heuristic extraction.
Modern LLM deployments suffer from escalating costs and latency. Traditional systems push entire histories into the model and hope relevance emerges.
Injecting full transcripts wastes tokens. We reduce prompt size by 60–90% by injecting structured summaries instead.
Bloated contexts lead to latency and "lost in the middle" phenomena. We enforce dynamic context ceilings.
Don't retrieve identical chunks repeatedly. Our Cost-Aware Invocation Engine calls LLMs only by exception.
| Layer | Purpose | LLM Required? |
|---|---|---|
| Ephemeral Context | Short-lived session memory | No |
| Working Memory | Current task state | Rare |
| Condensed Memory | Summarized structured knowledge | No |
| Indexed Semantic | Embedding searchable fragments | No |
| Strategic Memory | High-level distilled insights | Rare |
All conversational events, metadata, decisions, structured outputs, and tool results are captured.
We compress interactions into Intent summaries, Decision points, and State changes. This eliminates narrative redundancy.
Proof Envelopes wrap Assertions in a JSON envelope containing the hash of input events, ensuring auditability.
Sub-linear scaling with usage
This is not a vision piece. This is infrastructure.
We do not embed raw transcripts. We extract decoded Intent, Entities, Decisions, and Outcomes. Only condensed semantic units are stored.
We never re-summarize history. New interactions generate a semantic diff. This prevents recursive summary drift and exponential token growth.
"Neurons that fire together, wire together." Memories co-retrieved during successful agent actions form stronger semantic connections automatically.
Frequently accessed pathways are reinforced, moving critical knowledge from "transient" to "permanent" storage tiers.
Integrated TinyBERT NER models identify People, Organizations, and Locations in real-time, populating the Graph automatically.
Queries don't just hit the index. They trigger a "wave" of activation through the knowledge graph to uncover 2nd-order relationships.
Built on transparency and community-driven standards. Condensate is designed to be a foundational layer for the AI ecosystem.
Every assertion extracted from raw data passes through configurable guardrails before entering long-term memory. Humans stay in control of what the agent learns.
Multi-layer heuristic engine detects imperative commands, meta-instructions, and prompt injection patterns before they enter the knowledge graph.
Every assertion receives an instruction_score and safety_score (0.0–1.0). Configurable thresholds determine what
gets blocked vs. flagged for review.
Choose between manual review (all assertions require approval) or auto mode (guardrails only, no human step).
When AI agents ingest user-provided data, malicious actors can embed commands that corrupt long-term memory. Condensate's guardrails detect and block these before they ever reach the knowledge graph.
Condensate connects with your entire AI stack. Whether you're running local models or scaling on the cloud, we providing the memory layer.
import { CondensateClient } from '@condensate/sdk';
const client = new CondensateClient('http://localhost:8000');
// Add an Episodic Item
await client.addItem({
project_id: 'my-project-id',
source: 'chat',
text: 'The auth service fails on 500 error retries.',
metadata: { actor: 'AegentOS' }
});
// Retrieve Knowledge
const result = await client.retrieve({
query: "What are the issues with auth service?"
});
console.log(result.answer);import "github.com/condensate/condensate-go-sdk"
client := condensate.NewClient("http://localhost:8000", "api-key")
item := condensate.EpisodicItem{
ProjectID: "project-123",
Source: "incident",
Text: "Database latency spiked to 500ms",
}
id, err := client.AddItem(item)Join the standard for Agentic Memory. Spin up the full stack in seconds.
pip install condensate
npx -y @condensate/core
cargo add condensate