Retrieval depth. Or artifact horizon.

The honest read

Most "vs" pages exist to make the author look better. This one does not. Hindsight publishes openly, runs the same benchmarks we do, and is the closest peer in the memory-for-LLMs space. Two of their authors review benchmark methodology in public and we have learned from those threads.

We pair this with our own how-we-benchmark page so you can compare numbers head-to-head.

Where Hindsight wins

Honest concessions, in order of importance.

Retrieval depth. Their TEMPR fusion runs 4 distinct retrieval strategies and merges results. Our retrieval is single-strategy (FTS5 + embedding rerank). On long-context, multi-hop questions where the answer requires walking a temporal chain, they will out-recall us most days.

Memory model maturity. They have a more developed taxonomy of memory types (episodic vs semantic vs predicate) and the academic backing (LongMemEval co-leaderboard) to justify it. Our memory model is task-shaped, and we don't pretend it's a substitute for theirs on memory-only benchmarks.

Open community. Hindsight ships an open framework with active contributors. We ship a hosted compiler with an open spec (RS-1) but the trainer is closed-source. If you want to read the algorithm and tune it, Hindsight is the one to fork.

Where kolm wins

Consolidation, not just retrieval. The captured pairs aren't just for recall - they're labelled training data. We run them through a verifier, filter to the high-confidence subset, and distill a smaller model. Nine consolidation strategies feed the trainer; retrieval is one of them, not the whole thing.

The deliverable is a model, not a recall API. Hindsight gives you back better context to feed to a frontier model. We give you back a .kolm file - the model itself - signed, portable, runnable offline. Once compiled, you don't need the frontier for that task.

Receipts and signing. Every output a .kolm produces ships with an HMAC-SHA256 receipt chain. Hindsight has no equivalent because there's no artifact to sign.

Cost shape. Hindsight + frontier scales linearly with traffic; you pay for the LLM call every turn. kolm scales as a fixed compile cost + $0 marginal inference (the .kolm runs on your hardware). Past the compile threshold, the cost curves diverge.

Side-by-side

When to use Hindsight

Use Hindsight when the task requires retrieving the right memory at turn-time with high recall on multi-hop questions, and you want an open framework you can read and tune. The TEMPR fusion is the right tool when the question is "did the agent remember the relevant fact from three sessions ago?"

When to use kolm

Use kolm when the task requires a signed, portable model that runs without the frontier. Once a namespace has enough captured pairs, kolm consolidates them into a .kolm file you can ship with your product, run on a phone, or deploy on a customer's air-gapped network.

# point traffic at the kolm capture proxy:
ANTHROPIC_BASE_URL=https://kolm.ai/v1/capture/anthropic

# once enough pairs accumulate, compile:
kolm compile "answer support tickets" \
  --namespace support \
  --base qwen2.5-7b

ok wrote support.kolm  k_score=0.89  signature=hmac-sha256

Can I use both?

Yes, and the combination is well-formed. Hindsight on the inbound side for state-of-the-art recall feeding the frontier; kolm on the same captured pairs to accumulate a compileable corpus. Once a namespace hits the compile threshold, kolm replaces the frontier hop for that task. Hindsight keeps doing what it does best on the long-tail traffic.