Docklee Universal Memory

Your AI starts from
scratch every session. DUM fixes that.

AI applications are stateless by default. Every new session, your AI meets the user for the first time. DUM extracts facts from every interaction, stores them with confidence scores, resolves contradictions, and injects full context at the start of every session.

  • Extracts structured facts from conversation transcripts automatically
  • Stores with confidence scores, session source, and timestamps
  • Resolves contradictions when users update their stated context
  • Injects assembled memory in a single recall() call at session start
Without DUM
With DUM

An AI that forgets everything between sessions is not a product. It is a demo that runs on repeat. DUM gives your AI the memory layer it was never shipped with.

Without DUM

Session 3. Your AI asks a returning user what they do for work. They already answered this in Session 1 and Session 2. They close the tab.

With DUM

DUM recalled 'Senior Engineer at a 40-person fintech team'. Session 3 opened with that context. The AI picked up exactly where they left off.

Without DUM

A user said 'I prefer concise answers' in session 4. In session 9 the AI is verbose again. The preference was never stored anywhere.

With DUM

DUM stored the preference with confidence 0.96, flagged it as a strong signal. Every session after that inherits the user's stated preference.

Without DUM

User says they are a PM in session 1. In session 5 they say they moved to an engineering role. Your AI now contradicts itself mid-conversation.

With DUM

DUM detected the conflict, reconciled 'PM' and 'Engineering Lead' into one fact with merged confidence, and updated the context for all future sessions.

How DUM works

Write once.
Recalled forever.

Step 01

Pass DUM your conversation transcript

After each session ends, pass the raw transcript to DUM. It extracts structured facts, preferences, intent signals, and stated context automatically. You do not define the extraction schema. DUM figures out what matters.

Step 02

DUM stores, scores, and resolves

Every extracted fact is stored with a confidence score, a timestamp, and the session it came from. When facts contradict each other across sessions, DUM reconciles them into a single, merged representation rather than keeping conflicting versions.

Step 03

Call recall() at the start of every session

A single API call returns a structured context block with all high-confidence facts and a natural language summary ready to inject directly into your AI's system prompt. Your AI knows who it is talking to before the first message.

DUM · user_7x91k6 facts · 3 sessions · updated 47 min ago

Extracted facts

roleProduct Manager
99%
s1
companyFintech startup
97%
s1
team_size12 people
95%
s2
goalQ4 deployment
91%
s2
concernBudget approval
88%
s3
preferenceTechnical depth
84%
s3

Resolved conflict

s1: "Product Manager" vs s3: "Engineering Lead" · reconciled to "PM-turned-Engineering-Lead" · confidence 0.82

Memory lifecycle

Not just storage. A full memory pipeline.

DUM is not a key-value store for chat history. It runs a five-stage pipeline that extracts, scores, reconciles, decays, and assembles memory the way a person's memory actually works.

Facts pulled from every interaction

DUM reads each AI interaction and extracts structured facts: roles, preferences, goals, pain points, and stated context. Raw conversation text becomes a typed, queryable memory.

Where to use DUM

Any AI that talks to the same user more than once.

If your AI has returning users, it needs DUM. The pattern is always the same: stop starting from zero.

Support chatbots

A returning customer opens a support chat. Without DUM, the bot asks for their account details again. With DUM, it already knows their plan, their open ticket from last week, and that they prefer chat over email. The issue gets resolved in one exchange.

AI copilots

An in-product assistant powered by LLM has no idea who the user is unless you tell it every time. DUM lets your copilot say 'I see you're still building the analytics module' rather than 'How can I help you today?'

Onboarding flows

Onboarding flows ask the same qualifying questions on every return visit. DUM stores what the user told you in session 1 and uses it to skip ahead intelligently. The experience feels designed, not repeated.

Sales and CRM AI

Sales AI that knows a prospect mentioned a competitor, flagged a budget constraint, and is evaluating for Q4 will have a fundamentally different conversation than AI that starts cold. DUM is the memory that makes the difference.

What you get

Everything memory needs. Nothing you have to build.

Automatic fact extraction

Pass a conversation transcript. DUM extracts facts, preferences, and intent without you writing any extraction logic or defining a schema.

Confidence scoring

Every stored fact has a confidence score between 0 and 1. DUM uses extraction certainty, repetition frequency, and recency to keep scores accurate.

Contradiction resolution

When a user updates their context across sessions, DUM detects the conflict and reconciles both facts into a single merged representation with a new confidence score.

Memory decay

Facts that are never reconfirmed gradually decrease in confidence. Facts stated repeatedly across sessions strengthen. No manual expiry or cleanup required.

Per-user isolation

Every user's memory is completely isolated. recall() only returns facts for the exact user ID you pass. No cross-user leakage. Compliant with data isolation requirements.

Sub-50ms recall latency

The recall() call assembles and returns the full context block with typical latency under 50ms. Memory injection adds no meaningful latency to session startup.

Integration

Two calls.
write() and recall().

After each session: write the transcript. Before each session: recall the context. That is the entire integration. DUM handles everything in between.

  • No extraction schema to define. DUM figures out what matters.
  • Confidence scores and timestamps included on every recalled fact
  • Natural language summary ready to paste directly into a system prompt
  • Resolved contradictions surfaced explicitly so nothing is silently lost
Live
import { Dock, Lynk } from "docklee";
const dock = new Dock({ memorySpaceId: "space_app" });
const lynk = new Lynk({ memorySpaceId: "space_app" });
// After a conversation ends, capture it into DUM
dock.capture(conversationTranscript, {
type: "conversation",
endUser: "user_7x91k", // isolates memory per end user
// DUM handles extraction, scoring, and deduplication
});
await dock.flush();
// At the start of the NEXT session, recall what matters
const memories = await lynk.recall("who is this user and their goals", {
endUser: "user_7x91k",
topK: 5,
});
console.log(memories.map((m) => m.content));
// ["Engineering Lead at a 12-person fintech startup",
// "Targeting Q4 deployment", "Prefers technical depth"]
// Inject into your AI's system prompt
const systemPrompt = `You are a helpful assistant.
User context: ${memories.map((m) => m.content).join("; ")}`;

Docklee Universal Memory

Stop building AI that
forgets who it is talking to.

Add write() after each session. Add recall() before the next one. Your AI will never start from scratch again.