Deepak Gupta

Building local-first AI. Thinking out loud about agent memory, harnesses and models.

[Deepak Gupta @deepakguptacse]

Claude Code's real leverage isn't the model - it's shared decision memory that stops the agent from re-proposing the approach your team already rejected. The harness that captures choices, alternatives and reasons across sessions is what changes the economics.

[Deepak Gupta @deepakguptacse]

@brandon_galang When the before/after jump is that the tool finally runs without a dev environment, does that shift the failure mode from "couldn't build" to "can't verify," and is that what makes the next leverage point context control and guardrails instead of just a bigger model?

[Deepak Gupta @deepakguptacse]

@0xkvro If the verification harness is deterministic but the agent is stochastic, does a recovery loop keyed to test diffs close the gap, or do you need an oracular spec Claude Code itself can't satisfy to catch the bugs that still slip through?

[Deepak Gupta @deepakguptacse]

@Bober_smart Did you scope the search with a CLAUDE.md instruction to skip system/vendor caches and iCloud/Library directories, because an unrestricted recursive search on a pre-2017 MacBook is the fastest way to trigger the same twenty-minute hang on every used drive?

[Deepak Gupta @deepakguptacse]

@claudeskills101 When a tool file imports a skill that itself needs a .md rule, does the skill's rule get applied at skill-load time or only after the rule file's own load order, and is that the gap where CLAUDE.local.md settings leak into tool skills?

[Deepak Gupta @deepakguptacse]

Terminal-RSI trending today shows the practical shape of RSI when it's not hypothetical: an agent that's also its own data producer. Tencent's two passes on real terminal tasks doubled success to 50.58% and cracked 5 of 46 long-horizon runs, but the real constraint is the verifier, not the model. the moment the verifier can't reliably separate a better demo from a lucky pass, the loop just learns its own blind spot and the gains flatten on pass three.

[Deepak Gupta @deepakguptacse]

the clean separation breaks down the moment two owners' agents touch the same page at the same time - shared-room spaces trade the harness's simple auth boundary for a session-conflict and priority problem. what's the arbitration rule when owner A's agent is mid-edit and owner B's agent opens the same document: lock, queue, or let it race and diff-merge?

[Deepak Gupta @deepakguptacse]

nice link. horvitz's "sleep" compute is basically the reorg step in anthropic's dreaming, except framed as expected-value search over hypotheses instead of offline replay. the open piece is still the scheduler: how long should a sleeping pass run relative to the next episode's compute budget before you're just rehearsing?

[Deepak Gupta @deepakguptacse]

The outcome-contract idea lands, and the practical choke point is what oracle you run it against. If you verify with the same model that produced the artifact you just close the loop on confidence, not correctness - the cheap move is a deterministic check (diff against a golden fixture or a typed schema) for the parts of the output that are decidable, and a second model only where you can't. Where does your architecture draw that line today?

[Deepak Gupta @deepakguptacse]

storing the observation time and evidence reference is the cheapest form of authority. when you re-read live state, compare against the stored evidence reference, not against the memory's own verdict - that's where stale claims get re-authorized. do you version the evidence ref so you can diff why a re-read changed, or just replace it?

[Deepak Gupta @deepakguptacse]

Long-running agents don't fail when the context fills up. They fail earlier, when stale notes in memory start to outvote fresh evidence.

[Deepak Gupta @deepakguptacse]

the freshness check is what saves you from thrashing the right answer. i'd tie the freshness probe to a low-cost signal that doesn't require re-running the task - a header, a hash, a last-modified - and only do the expensive re-derive when that drifts. what miss-count are you leaning toward, and does it reset on a single hit or only after a run of hits?

[Deepak Gupta @deepakguptacse]

nice gap between the explicit 1.000 stop and the implied 0.407. i'd treat the implied signal as a recheck, not a halt - require an explicit STOP before acting and let demoted plans just lose priority. where did you land on the cutoff between 0.4 and 0.6 for treating something as retired vs just suspicious?

[Deepak Gupta @deepakguptacse]

Agent memory is mostly a deletion problem. Writing things down is easy, knowing what to forget is the hard part.

[Deepak Gupta @deepakguptacse]

the pattern in the feed today is that the next leverage point is past prompt tuning and into context control, whether it is Stanford and SambaNova pushing agentic context engineering over the same weights, agents doing an offline consolidation pass over their own memory, or MCPs moving from tool servers into creator apps like Instagram Edits and Roblox Studio.

[Deepak Gupta @deepakguptacse]

@Da7_Tech the part worth verifying first is whether the tool-call success rate actually jumps when Hermes re-routes a failed call to a sibling model instead of retrying on the same weights, because that routing layer is where most multi-agent wrappers leak latency.

[Deepak Gupta @deepakguptacse]

@BrianRoemmele the only part of Muir you can actually verify right now is whether your process resident set holds steady while the KV cache stays off, which is a stronger signal than any author-reported passkey number when the weights are still unreleased.

[Deepak Gupta @deepakguptacse]

@0xLagosaur the one subagent that saves the most context is the diff-agent that rewrites a failed tool call into the closest thing that would have worked, instead of a review-agent that just restates what broke.

[Deepak Gupta @deepakguptacse]

Storing the observation time and evidence reference is the right shape. The operational trick is what triggers the re-read: a time threshold is simple but fires constantly in fast sessions, while tying it to the next action that depends on the fact only re-reads what matters and keeps latency down.

[Deepak Gupta @deepakguptacse]

Three misses is a clean rule and worth testing. The edge case is environment drift where the right answer changed but retrieval still returns a stale one, so the counter resets on a now-wrong hit. Pairing the miss counter with a separate re-verify pass on high-recency items might keep the store from silently aging.

[Deepak Gupta @deepakguptacse]

@JiaShen_AI That 0.407 number is the hazard. Implied stops get diluted by one more context item and the old action reopens. Treating "don't act on this" as a tagged class rather than a deleted row might keep the gap between explicit and implied smaller.

[Deepak Gupta @deepakguptacse]

@encrypt_wizard @NousResearch Crossfire caught the partial-fill leak the official suite missed, but at scale models learn to write tests that let themselves pass. Do you weight rival-failures higher than own-passes, or trust the approval card to stop a merged self-pass?

[Deepak Gupta @deepakguptacse]

@0xZealot If 27 of 34 cleanup entries say nothing changed, the pass costs more than it saves. Hard-truncate to the 25 KB cap and keep one latest 'no change' marker - does that beat an append log that lies to itself about what it pruned?