Skip to content

Latest commit

 

History

History
214 lines (174 loc) · 10.1 KB

File metadata and controls

214 lines (174 loc) · 10.1 KB

Magic Moments, Claim-Bounded

State: current public receipt note; first useful shape leads before caveats.

First Useful Shape

old cue -> source-backed snippet or reopenable route -> source boundary -> next action

Example public-safe receipt:

  • Cue: "without pretending it has innate memory"
  • Source-backed hit: examples/public-memory-bundle/clean-source, message msg_public_001.
  • Snippet: the old user text asks whether an agent can catch up without pretending it has innate memory.
  • Boundary: this supports the source trail and exact quote only after source is reopened; it does not prove broad recall quality or innate model memory.
  • Next action: deepen/open the source turn before using exact wording or turning the route into a public claim.

Role: product and human evidence.

Status: current claim-bounded live-use examples; not benchmark scores, release claims, or proof of innate model memory.

AIppocampus should not make a new reader dig through benchmark caveats before seeing why the project exists. This page collects a few real second-user moments where source-backed continuity felt different from ordinary chat memory, while keeping the claim boundary visible.

These examples come from external live-use notes in Discussion #98 and the reader-facing field report in Discussion #428. They are product-shaped evidence, not benchmark scores, release claims, or proof that the base model remembered anything by itself.

How To Read These

  • A memory scent or hook result is navigation, not evidence.
  • Specific claims should reopen clean source, registry state, automation state, or another durable source before they are trusted.
  • Local paths, session ids, credentials, raw private snippets, and unnecessary personal detail are intentionally omitted.
  • The examples show useful continuity moments; they do not prove universal recall quality, no-hook superiority, hosted-service readiness, or private real-history benchmark coverage.

Examples

Fresh Projectless Thread

Source: fresh-thread live-use note.

  • User typed: "do u know what i'm working on recently?"
  • AIppocampus helped recover: several recent work streams, including a phonics app, a signal-scanner product, and AIppocampus/OpenClaw-adjacent automation work, with uncertainty marked instead of flattened away.
  • What made it source-backed: the note says the workspace had no relevant project files and no project-local instructions containing those facts. A light ambient scent oriented the agent, then the agent reopened local memory/registry evidence and direct clean source before making specific claims.
  • What not to claim: this was not a no-hook baseline and does not prove universal fresh-thread recall quality.
  • Why it matters: a new, projectless thread did not start from bare ground, but it still treated scent as a cue rather than as proof.

Correction Across Languages

Source: fresh-thread live-use note.

  • User typed: a Russian question about which words had been studied, then corrected the route: "нет я не про сайт. про школу".
  • AIppocampus helped recover: the first answer routed to the phonics app context; after the correction, the assistant moved to the school/Oxford Phonics context and gave a narrower, uncertainty-marked answer.
  • What made it source-backed: the useful behavior was not that the first route was perfect. It was the progressive route change after a small user correction, while preserving the difference between app/project context and school context.
  • What not to claim: this does not prove the first route is always right or that multilingual recall is solved.
  • Why it matters: long-running continuity needs recoverable correction, not only confident first guesses.

Ambiguous LinkedIn Cue

Source: fresh-thread live-use note.

  • User typed: "что случилось с моим линкедин?"
  • AIppocampus helped recover: a local automation/delivery failure around LinkedIn draft delivery, while refusing to claim live LinkedIn account state.
  • What made it source-backed: the assistant separated external account state from local automation evidence, inspected local automation state, and framed the live-account part as unverified.
  • What not to claim: this does not prove browser/live-account access or that anything was verified on LinkedIn itself.
  • Why it matters: source-backed memory should prevent overclaiming as much as it helps recall.

Long-Thread Fuzzy Self-Reference

Source: long-thread live-use note.

  • User typed: a fuzzy question asking what an earlier "xxx" completion question had referred to.
  • AIppocampus helped recover: the target was Phonics Lab Books 3/4/5, with the important nuance that image assets were effectively complete while audio/content/product-course closure was not fully complete.
  • What made it source-backed: the thread was roughly multi-day and 140+ visible user-message events long. The hook gave general orientation, but it did not directly surface the decisive anchors. The agent still had to inspect rollout metadata, use clean/turn tooling, and search clean source/raw text.
  • What not to claim: this does not prove that the foreground hook alone is strong enough, nor does it replace private real-history benchmark tracks.
  • Why it matters: this is the lived product promise: fuzzy old references can be recovered when the agent is allowed to reopen source instead of pretending to remember.

Cross-Thread Tool-Failure Provenance

Source: second-user field report.

  • User typed: in a later thread, the user asked what had been asked in a separate fresh thread and what went wrong during tool work.
  • AIppocampus helped recover: the earlier prompt shape across a language boundary, the distinction between two Discord workflow-error families, and a concrete Python tool-failure kind from the earlier investigation.
  • What made it source-backed: the assistant relied on registered clean source and raw audit evidence rather than treating the follow-up as something the base model remembered. The public benchmark fixture records only the scenario shape and sanitized hashes, not the raw prompt or raw error text.
  • What not to claim: this does not prove ambient-hook-only exact recall, universal cross-thread recovery, or that the tool failure was caused by the memory system.
  • Why it matters: continuity should preserve mundane agent-work provenance, including mistakes and recoveries, without turning those details into private public fixtures.

Try A Public-Safe Path

For the canonical first-recall flow, use docs/guides/first-recall-decision-card.md.

For Codex Desktop users who asked an agent to set up AIppocampus, start with the trusted local plugin path and then ask for one foreground continuity route:

aippocampus plugin install --codex --verify
aippocampus update status
aippocampus agent recall "old decision or handoff cue" --json
aippocampus agent deepen --request 1 --recall-selector <emitted-selector> --json

The no-clone public-safe probe remains useful when the user is evaluating the package without private history:

uvx aippocampus --help
uvx aippocampus onboard --provider auto --status

After explicit consent to register selected local history, ask for a source-backed continuity route. Exact search is the fallback/demo path when the user remembers wording:

uvx aippocampus onboard --provider codex --status --json
# Then follow the explicit write recommendation after consent.
uvx aippocampus agent recall "old decision or handoff cue" --json
uvx aippocampus agent deepen --request 1 --recall-selector <emitted-selector> --json
uvx aippocampus search "a distinctive old phrase"

If the user does not remember exact wording, use a project cue or time cue as candidate navigation only. Do not present a vague-cue route as evidence until AIppocampus returns a source-backed snippet.

PyPI and MCP Registry publication evidence is captured in #291. Broader all-client readiness claims still require the separate external install/UI readiness track in #307.

For no-private-data demos, start with docs/guides/demo-scenarios.md. For the benchmark and smoke ledger behind broader claims, use benchmark-evidence-map.md and readiness/stage-0-5-readiness.md.

The claim-bounded benchmark slice for turning these reports into reproducible fixture coverage is #454, with its public fixture report in benchmarks/reports/field-journey/field-continuity-fixture-report.md. The broader Field Continuity Eval design for public reproducibility tracks, baselines, metrics, and private-dogfood boundaries is benchmarks/field-continuity-eval-design.md.

Related technical tracks: #201, #281, #285, #291, #382, and #397.