Cognitive infrastructure for AI agents
Understand context.
Learn from experience.
DeepDream researches how AI agents interpret business context, update long-term memory, and apply past lessons to the next task.
- Semantic understanding
- Long-term memory
- Cognitive intelligence
Technical architecture
From context to reusable experience.
Understand the concepts and constraints first. Then decide what to retain, when to recall it, and how to verify that it helps.
Semantic understanding
Resolve what entities, terms, and constraints mean for a specific team at a specific time, beyond keyword matching.
Long-term memory
Write selectively, update over time, and recall for the task. Memories need to be traceable, correctable, and forgettable.
Cognitive intelligence
Use task feedback to verify lessons, extract reusable strategies, and test whether they transfer to new tasks.
FEATURES / DREAM
Make the last task useful for the next.
Dream is DeepDream’s design for background memory maintenance. Between sessions, it reviews task logs, tool results, and feedback to update the memory an agent can read without changing model weights.
- Startup context
- Confirm the team and metric version before analyzing.
- Knowledge representation
- A metric page links both definitions, the change record, and query templates.
- Memory notes
- The growth team changed its criterion from login to a core action.
- Raw traces
- Keep the approval record and query results for later verification.
- CORRECT
Check memory against the evidence
When a note conflicts with logs or verified results, distinguish an original mistake from a later rule change. Correct the mistake, retain dated versions for rule changes, and flag uncertain notes for review.
- ORGANIZE
Give scattered notes a structure
Merge duplicates by entity or topic, then connect related notes through pages, indexes, and links. Keep source records searchable so an agent can check the original, not just the summary.
- PRIORITIZE
Preload only what matters
Keep a short startup context with verified, frequently useful constraints and pointers. Load details on demand and move stale items out of startup context.
- FOCUS
Keep what serves the task
For metric analysis, prioritize definitions, effective dates, and validation steps. Use task feedback to decide which rules are worth keeping rather than saving every action as lasting memory.
- FILL GAPS
Find what one session missed
Compare sessions to uncover recurring mistakes and missing conditions. Add notes backed by evidence, repair broken links, and leave unsupported gaps unresolved.
Use case · DATA AGENT
When “active user” changes meaning
A growth team revises its definition of active users. When the next anomaly appears, the agent needs to know when and why the definition changed, and which analysis steps still apply.
Illustrative scenario · not production data
Metric: weekly active users (WAU)
Resolve the metric definition before running the analysis
Different teams can use the same metric name for different definitions. Identify the team and reporting period. If the definition is still ambiguous, ask rather than assuming the last one applies.
Team and time rangeConfirmed change: login → core action
Record both versions and when each applies
Record the revised definition, its source of approval, and its effective date. Keep the old version for historical queries without treating it as current.
Evidence and effective datesRecalled: a verified channel breakdown
Apply only the lessons that still hold
Check task fit, access, timing, and provenance. Keep the channel breakdown method, but reject the SQL template that still filters by the old definition.
Relevance and scope of useCheck: recompute the trend with one definition
Separate a definition change from a real business change
Recompute comparable periods with a consistent definition and record whether the evidence supports the conclusion. Save useful steps with their conditions so they can be tested again.
Outcome and feedbackDEEPDREAM RESEARCH
Remembering is not enough.
Explore how semantic modeling, the memory lifecycle, and feedback help agents use past experience on real tasks.