Discover
Find the earlier investigation by problem or file.
Case file TK-01Illustrative workflowSource-backed recall
The invisible repository
Turning AI session history into codebase memory.
How a developer can recover the reasoning behind an unfamiliar file without repeating the original investigation.
01
Outcome
Explain
Recover who changed it, what was tried, and why.
Continue
Load only the evidence needed for the current task.
This demonstrates a supported workflow. It does not claim a measured reduction in onboarding time, incident duration, or engineering cost.
02
Incident timeline
The second bug report looks familiar. The engineer does not.
Six weeks after an authentication regression was resolved, it returns in a related flow. The code contains the fix, but not the experiments, rejected approaches, or operational constraint behind it.
01
Day 0
Investigation
An engineer and coding assistant trace the failure, inspect logs, reject two hypotheses, change a validation boundary, and add a regression test.
02
Week 6
Recurrence
A different engineer sees similar symptoms. Git shows the patch, but not why the other approaches failed.
03
Recall
Recovery
The current assistant finds the relevant session and produces a focused brief linked to its source.
03
Knowledge map
Existing systems remember fragments. The session preserves the investigation.
ReclaimLLM complements the codebase's sources of truth by preserving the conversational and tool-use trail they usually omit.
SourceWhat it remembersWhat is still missing
01
Code
The current behavior
Rejected approaches and the constraints that shaped it
02
Git
The diff, author, and timestamp
The investigation that led to the change
03
Tickets
The requested outcome
The detailed reasoning produced while doing the work
04
AI sessions
Questions, tool activity, decisions, and file changes
Anything never captured or private to the provider runtime
04
Recall protocol
Search broadly. Inspect narrowly. Load context deliberately.
Begin with lightweight discovery and expand only after selecting a relevant source. Recall stays useful, auditable, and proportional to the current task.
01search_sessions
Find the investigation
Search by intent and project, even when the second engineer describes the problem differently.
> Find the previous investigation into intermittent auth callbacks.02file_brief
Trace the affected file
See which sessions touched the file, who did the work, when it happened, and why.
> Give me a history brief for server/auth/token_service.py.03get_session
Inspect the evidence
Open the selected source session before relying on a generated explanation.
> Show me the source session before I rely on this decision.04summarize_session
Bring forward only what helps
Load a focused summary or explicitly transfer the complete captured record when needed.
> Summarize this session as context for the change I am making now.05
Evidence packet
A briefing that shows its work.
The useful artifact is not a vague answer that the system “remembers.” It is a concise account of prior work with a source session the engineer can inspect.
Illustrative content / not a production record
FILE BRIEF / token_service.pySource linked
- What happened
- The earlier session isolated the failure to validation order after rejecting two competing hypotheses.
- Decision
- Preserve signature verification before normalizing the incoming payload.
- Files
- server/auth/token_service.py and its regression test
- Open work
- Confirm the adjacent callback path follows the same invariant.
- Source
- Session metadata, summary, tool activity, and recorded file changes
06
Truth boundary
Supported
- Search captured sessions by intent, project, path, or metadata.
- Recover file history with author, timing, and session context.
- Reuse summaries, relevant diffs, and pending work.
- Apply ownership and organization visibility rules to retrieval.
- Keep historical commands read-only unless new work is explicitly started.
Not supported
- Know about work that was never captured.
- Guarantee an old decision is still correct for the current code.
- Replace ADRs, documentation, review, or engineering judgment.
- Reconstruct hidden prompts, permissions, caches, or runtime state.
- Bypass workspace, organization, or encrypted-session controls.
07
The operating model
Tribal knowledge becomes useful when it is captured, searchable, and attributable.
ReclaimLLM does not make old reasoning automatically correct. It makes the evidence discoverable so the next engineer can validate it against the current code instead of beginning from zero.