Labeeb
Holographic vs Hindsight Memory
Everything the agent is thinking about, packed into one compact snapshot. Ask by role — the file, the function, the constraint — and pull the value back out.
- Strong for dense, fuzzy recall during deep work
- Weak at temporal reasoning; failures blur into the field
Episodes retained, recalled across semantic + keyword + graph + temporal strategies, and reflected into consolidated observations with trend signals.
- Strong for cross-episode judgement and pattern recognition
- Weak at live, high-speed scratchpad work
Why would an agent team need more than one kind of memory? The Studio runs two — Holographic and Hindsight — and assigns them to different roles. They’re built from different primitives because they answer different questions.
Holographic memory
“Holographic” here isn’t metaphor — it points at a real technique called Holographic Reduced Representations. The idea: everything the agent is currently thinking about — the file it opened, the function it called, the constraint it was handed — gets packed into a single compact representation that doesn’t grow as you add to it. Think of it as working memory squeezed into a fixed-size snapshot.
To look something up, the agent asks by role rather than by address. Not “give me record 42,” but “given this file, what’s the function I was editing?” — and the answer comes back out of the snapshot. A partial or fuzzy question still returns a plausible answer, and the lookup is cheap enough to run constantly in the inner loop.
- Fits. Dense, interconnected working context. Holding a whole surface in mind while editing it. Fuzzy lookups that still land on something usable.
- Breaks. Pack too much in and every recall blurs. No temporal ordering, no notion of “last time.” Failures get absorbed into the snapshot rather than remembered as failures.
- Runs under. Every engineer agent outside the planning layer — frontend, backend, AI engineer, QA, platform, database.
Hindsight memory
Hindsight is a different shape entirely. Built by Vectorize as a structured memory layer for agents, it stores information through retain, searches through recall, and rolls episodes into higher-level understanding through reflect. Reads run four retrieval strategies in parallel — semantic, keyword, graph, and temporal — so a question like “what did we decide last spring about the ingest pipeline?” doesn’t have to pick a single mode up front.
What’s interesting is the hierarchy underneath. Observations auto-consolidate overlapping facts into persistent beliefs, keeping an evidence trail and a computed trend — is this belief strengthening, weakening, going stale? Mental Models sit above those as curated summaries. The whole structure is temporal and relational by design, so “what happened the last time we tried this?” is a first-class query, not an afterthought.
- Fits. Reasoning across many episodes. Noticing that a plan looks like one that failed two weeks ago. Keeping outcomes attached to the decisions that produced them.
- Breaks. It lags live context — it’s a reflection layer, not a scratchpad. More bookkeeping too: episodes need boundaries, outcomes need labels.
- Runs under. The CTO agent and the principal engineers — the planners. Anyone whose job is to choose between paths rather than walk down one.
Why the split runs this way
The real question is: what’s the unit of thought you’re asking the agent to hold? For a builder — the frontend engineer halfway through a diff — the unit is a dense, tangled snapshot of the surface they’re editing. Holographic compresses that into something you can carry all at once and query algebraically. Does it forget its mistakes? Yes — but the builder isn’t being asked to remember them. They’re being asked to ship this change, carefully.
For a planner — the CTO deciding where to route work, a principal engineer weighing two designs — the unit is a past episode. Did a routing like this work last time? Did this design choice cost us later? Hindsight is built to answer that shape of question, and its observation-consolidation loop is what turns a stream of episodes into judgement.
Could one memory system do both? Probably, eventually. But trying to cram episode-level reasoning into a single working snapshot blurs the episodes, and trying to run deep working context through a multi-strategy retrieval hierarchy loses the shape of the surface you’re actually editing. For now it’s simpler to let each system do what it’s designed for, and pick the one that matches the job.