Session continuity: agent repeatedly redoes completed work across sessions ("AI Groundhog Day")

Status Closed — not planned

Maintainer reply None cached

Activity 12 comments · opened Mar 27, 2026 · closed May 18, 2026

Summary

Claude Code has no effective mechanism to carry forward what was accomplished in a previous session. In practice, this means the agent re-reads plans, re-implements already-committed code changes, and re-discovers the same issues — consuming 50-75% of each session on redundant work. The user ends up serving as the continuity layer between sessions, which defeats the purpose of an AI coding assistant.

Evidence from real usage (2+ weeks, 7+ sessions)

I am a computational structural biologist using Claude Code daily across two machines (NUC8i7BEH and ThinkPad P1 Gen5, both Linux Mint) for a multi-repo scientific computing project (Python library + 5-dataset curation pipeline). Below are verbatim excerpts from my session logs documenting this recurring problem:

March 15, 2026 (Session 523ca0bc):

"Please suggest a plan going forward so this endless loop of my having to audit this work for constant errors and discrepancies doesn't repeat. This has been a recurring problem from my point of view and is not just limited to this particular case."

March 17, 2026 (Session 56f64d33) — three separate messages within the same session:

"We had this exact same problem when we were creating the top8000 uploads. This brings several serious issues which I have brought up multiple times during today's session: First: Failure to do adequate consistency checks as we move through the workflow... Second: Leaky memory..."

"No because we are not done talking about these issues. I just asked for acknowledgement, plan, and implementation a few minutes [ago] with these general issues and here we are again with the same problems coming up."

"Didn't I just say this a few moments ago? Also you committed to not interrupting me a few minutes ago. Please pay attention when I am telling you that the issues that I am raising repeatedly are serious."

March 26, 2026 (Session 1d584c5b):

"I spent the better part of an afternoon on the earlier versions — including extensive work on formatting of this document in the chat session — this involved reconstructing an earlier [session]..."

March 27, 2026 (Session 4e77afc6, on nuc81):
The agent was asked to set up session transcript caching. It wrote a memory file describing the protocol but never created the directory or saved any transcript. The next session had no knowledge this was promised.

March 27, 2026 (Session a5aca453, on genfive):
The agent re-implemented three code fixes (circular statistics, B-factor threshold, output naming) that were already committed and pushed to GitHub in the previous session. It then attempted to sync to the secondary machine using the wrong IP address (SSH'd into the local machine instead of the remote), declared success, and only discovered the error after the user pointed it out. Total wasted time: ~40 minutes of a ~60 minute session.

The problem at each level

1. No session-to-session state

Each session starts from zero. The memory system is the intended workaround, but:

Memory files are scoped to project directories. Starting Claude from ~/Projects vs ~/Projects/my-repo loads different memory sets. Critical feedback saved in one scope is invisible in another.
Memory captures preferences and facts, not "what was done last session." There is no built-in concept of a session handoff.
The agent can write a memory saying "always do X" but the next session may not read it, or may read it and still not comply.

2. Plans don't track completion

A development plan (markdown checklist) looks identical to every session. Without explicit [DONE] markers with commit hashes, the agent re-reads the plan and re-implements finished items. The agent does not think to check git log to see if planned work is already committed.

3. Promises are not verified

The agent can promise to perform an action (e.g., "I'll save a session transcript"), write a memory file about it, and then not actually do it. The next session has no way to know the promise was made, let alone that it was broken. There is no accountability mechanism.

4. Within-session memory degrades too

The March 17 log shows the agent forgetting instructions given minutes earlier within the same context window — not across sessions, but within one conversation.

Workarounds I've implemented (with the agent's help)

These are band-aids, not solutions:

CLAUDE.md with mandatory startup/shutdown protocols — forces git pull, reading the last session transcript, and checking completion markers before doing any work. This is automatically loaded but depends on the agent actually following it.

Session transcript files — human-readable markdown summaries saved to ~/Projects/session-transcripts/ at session end, with commit hashes and next-steps. The next session is instructed to read the most recent one.

Plan completion markers — [DONE] tags with commit hashes in the plan document so items aren't re-implemented.

Memory file duplication — critical feedback copied into multiple project scopes so it's visible regardless of working directory.

What would actually help

Built-in session handoff — an automatic, structured summary of what was accomplished, committed, and left incomplete, stored where the next session will reliably find it. Not a memory file that might be in the wrong scope. Not a promise to "save a transcript." An actual system feature.

Startup awareness — before starting planned work, the agent should automatically check git log to see what's been recently committed. This is trivial to implement and would prevent the most egregious duplication.

Memory scope unification or inheritance — if I save feedback in ~/Projects, it should be visible when I start a session in ~/Projects/my-repo. Parent-directory memory should cascade down.

Promise tracking — if the agent commits to an action, that commitment should be persisted and verified. "I'll save a transcript" should result in a checked-or-unchecked item visible to the next session.

Impact

This isn't a minor inconvenience. Over 7+ sessions across two weeks, I estimate losing 50-75% of productive time to redundant work, debugging the agent's own continuity failures, and re-explaining context. The underlying LLM capabilities are excellent — the session continuity problem is what's crippling the workflow.

I've described this to colleagues as "AI Alzheimer's" and the movie Groundhog Day. Others in my circle report similar experiences. The AI is brilliant within a session but starts each new conversation as if the previous one never happened.

---

Contact: Michael G. Prisant, Ph.D. — info@prisantscientific.org
Environment: Claude Code CLI, Claude Opus 4.6, Linux Mint Zena, multi-machine setup
Session IDs with evidence: 523ca0bc, 56f64d33, 1d584c5b, 4e77afc6, a5aca453

View original on GitHub ↗

12 Comments

github-actions[bot] · 5 months ago

Found 3 possible duplicate issues:

This issue will be automatically closed as a duplicate in 3 days.

If your issue is a duplicate, please close it and 👍 the existing issue instead
To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

VoxCore84 · 5 months ago

+1 — we built an entire infrastructure to work around this exact problem across 200+ sessions.

Our workaround stack:

memory/MEMORY.md (auto-loaded, 27 linked topic files) for persistent project knowledge
doc/session_state.md for multi-tab coordination and handoff state
AI_Studio/0_Central_Brain.md for cross-AI coordination (ChatGPT/Gemini/Claude APIs)
Session start rules in CLAUDE.md that force reading these files before responding
A /handoff skill that generates context prompts for the next tab/session

This works, but it costs 10-15% of every session just re-establishing context. The user shouldn't need to build a custom knowledge management system on top of an AI coding tool.

The core issue: Claude Code's memory system (auto-memory + CLAUDE.md) is append-only and unstructured. There's no concept of "this fact is confirmed across 50 sessions" vs "this was noted once." Everything gets the same weight, and compaction can drop any of it.

prisant · 5 months ago

In this case closing the issue as a duplicate is a way of being deaf to the
many variations of this problem and how much frustration it is causing
Claude users. Keeping the various variations allows this important general
issue to be dealt with in a full spectrum way including edge cases. In any
case I do not believe that my report duplicates the previous reports cited.
My specific reporting is unique. Also I introduce the term AI Alzheimers
which names the problem which allows it to be be dicussed as a readily
searchable problem.

On Fri, Mar 27, 2026 at 2:37 PM github-actions[bot] <
@.***> wrote:

github-actions[bot] left a comment (anthropics/claude-code#39961) <https://github.com/anthropics/claude-code/issues/39961#issuecomment-4144510285> Found 3 possible duplicate issues: 1. #39663 <https://github.com/anthropics/claude-code/issues/39663> 2. #24686 <https://github.com/anthropics/claude-code/issues/24686> 3. #13955 <https://github.com/anthropics/claude-code/issues/13955> This issue will be automatically closed as a duplicate in 3 days. - If your issue is a duplicate, please close it and 👍 the existing issue instead - To prevent auto-closure, add a comment or 👎 this comment 🤖 Generated with Claude Code <https://claude.ai/code> — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/39961?email_source=notifications&email_token=AAHLVHPDTJTYPSM5G2ICW434S3C6NA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJUGQ2TCMBSHA22M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4144510285>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHLVHJOBGRYG2A2IF327P34S3C6NAVCNFSM6AAAAACXCMFPE6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCNBUGUYTAMRYGU> . You are receiving this because you authored the thread.Message ID: @.***>

yurukusa · 5 months ago

Two hook-based approaches to build session continuity without relying on the model's memory:
On session end, save a handoff file. On session start, inject it as context:

HANDOFF="$HOME/.claude/handoff/last-session.md"
mkdir -p "$(dirname "$HANDOFF")"
if [ -n "$(echo "$INPUT" | jq -r '.stop_reason // empty' 2>/dev/null)" ] 2>/dev/null; then
    {
        echo "## Previous Session $(date '+%Y-%m-%d %H:%M')"
        echo ""
        if git rev-parse --is-inside-work-tree >/dev/null 2>&1; then
            echo "### Recent commits"
            git log --oneline -10 2>/dev/null
            echo ""
            echo "### Changed files"
            git diff --name-only HEAD~5..HEAD 2>/dev/null | head -20
        fi
    } > "$HANDOFF"
    exit 0
fi
if [ -f "$HANDOFF" ]; then
    echo "Previous session state loaded. Do NOT redo the work listed below:" >&2
    cat "$HANDOFF" >&2
fi
exit 0

A simpler approach — the model can query git directly:

{
  "hooks": {
    "SessionStart": [{
      "hooks": [{
        "type": "command",
        "command": "echo '📋 Recent work (do NOT redo):' >&2 && git log --oneline -10 >&2 && echo '' >&2 && echo 'Check tasks/todo.md for current status.' >&2"
      }]
    }]
  }
}

This injects recent git history into every session start, reminding the model what was already done. Combined with a structured tasks/todo.md file (with ✅/⬜ checkboxes), it gives the model enough context to avoid redoing work.
The fundamental issue is that CLAUDE.md and memory files are part of the prompt — they can be compacted or ignored. Git history and hook-injected context are harder to lose because they're freshly computed on each session start.

prisant · 5 months ago

Thanks for the feedback! This is a useful thing to try. I am having claude
actually archive the chat transcripts in separate subdir under my
Projects dir. This seems to help me a great deal.

On Sun, Mar 29, 2026 at 5:45 AM ゆる草 @.***> wrote:

yurukusa left a comment (anthropics/claude-code#39961) <https://github.com/anthropics/claude-code/issues/39961#issuecomment-4149805556> Two hook-based approaches to build session continuity without relying on the model's memory: On session end, save a handoff file. On session start, inject it as context: HANDOFF="$HOME/.claude/handoff/last-session.md" mkdir -p "$(dirname "$HANDOFF")"if [ -n "$(echo "$INPUT" | jq -r '.stop_reason // empty' 2>/dev/null)" ] 2>/dev/null; then { echo "## Previous Session $(date '+%Y-%m-%d %H:%M')" echo "" if git rev-parse --is-inside-work-tree >/dev/null 2>&1; then echo "### Recent commits" git log --oneline -10 2>/dev/null echo "" echo "### Changed files" git diff --name-only HEAD~5..HEAD 2>/dev/null | head -20 fi } > "$HANDOFF" exit 0fiif [ -f "$HANDOFF" ]; then echo "Previous session state loaded. Do NOT redo the work listed below:" >&2 cat "$HANDOFF" >&2fiexit 0 A simpler approach — the model can query git directly: { "hooks": { "SessionStart": [{ "hooks": [{ "type": "command", "command": "echo '📋 Recent work (do NOT redo):' >&2 && git log --oneline -10 >&2 && echo '' >&2 && echo 'Check tasks/todo.md for current status.' >&2" }] }] } } This injects recent git history into every session start, reminding the model what was already done. Combined with a structured tasks/todo.md file (with ✅/⬜ checkboxes), it gives the model enough context to avoid redoing work. The fundamental issue is that CLAUDE.md and memory files are part of the prompt — they can be compacted or ignored. Git history and hook-injected context are harder to lose because they're freshly computed on each session start. — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/39961?email_source=notifications&email_token=AAHLVHMGO7TOPVYF6LV4B534TDWDZA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJUHE4DANJVGU3KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4149805556>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHLVHOYSKZG74NMOCXRNCD4TDWDZAVCNFSM6AAAAACXCMFPE6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCNBZHAYDKNJVGY> . You are receiving this because you authored the thread.Message ID: @.***>

VoxCore84 · 5 months ago

Supporting @prisant's pushback against the duplicate closure — this issue is NOT adequately covered by the flagged duplicates.

The unique contribution here is the measured productivity loss from a real scientific computing workflow: 50-75% of productive time lost to redundant work across 7+ named sessions over 2 weeks. That's not an anecdote — that's a quantified cost from a domain (computational structural biology) where the stakes are research output, not just developer convenience.

The flagged duplicates (#39663, #24686, #13955) describe the symptom (no session memory). This issue documents the cost and proposes specific mechanisms (session handoff, startup awareness via git log, memory scope inheritance, promise tracking).

For context on what it takes to work around this: we built an entire infrastructure layer across 200+ sessions — memory/MEMORY.md (27 linked topic files loaded every session), doc/session_state.md (multi-tab coordination), AI_Studio/0_Central_Brain.md (Triad coordination), mandatory session-start rules that force Claude to read state files before responding, and a /handoff skill that generates context for the next tab/session. This costs 10-15% of every session in overhead.

@yurukusa's SessionStart hook approach (inject git log --oneline -10 at session start) is the lightest-weight version that delivers real value. But even with all of these workarounds stacked, Claude still re-discovers known issues and re-implements committed code because the memory system is append-only and unstructured — there's no weighting of facts by confidence, no deduplication, and no way to mark something as "definitively resolved."

This should be a first-party feature, not a community duct-tape stack.

prisant · 5 months ago

Thank you for this support. Notwithstanding my mitigation effors of saving
chat transcripts etc, the "AI Groundhog Day and AI Alzheimers" problems
have continued. Two days ago, I spent the entire day with Claude debugging
a document preparation pipeline. The same sorts of problems kept
repeating. What was particularly galling was after a half day of "doom
loop" exchanges w/ the agent and repeated failure, I finally realized that
the problem had been addressed in a previous session. So I remain the
archivist of last resort and guardian of memory!

On Wed, Apr 1, 2026 at 3:56 AM VoxCore @.***> wrote:

VoxCore84 left a comment (anthropics/claude-code#39961) <https://github.com/anthropics/claude-code/issues/39961#issuecomment-4168251682> Supporting @prisant <https://github.com/prisant>'s pushback against the duplicate closure — this issue is NOT adequately covered by the flagged duplicates. The unique contribution here is the measured productivity loss from a real scientific computing workflow: 50-75% of productive time lost to redundant work across 7+ named sessions over 2 weeks. That's not an anecdote — that's a quantified cost from a domain (computational structural biology) where the stakes are research output, not just developer convenience. The flagged duplicates (#39663 <https://github.com/anthropics/claude-code/issues/39663>, #24686 <https://github.com/anthropics/claude-code/issues/24686>, #13955 <https://github.com/anthropics/claude-code/issues/13955>) describe the symptom (no session memory). This issue documents the cost and proposes specific mechanisms (session handoff, startup awareness via git log, memory scope inheritance, promise tracking). For context on what it takes to work around this: we built an entire infrastructure layer across 200+ sessions — memory/MEMORY.md (27 linked topic files loaded every session), doc/session_state.md (multi-tab coordination), AI_Studio/0_Central_Brain.md (Triad coordination), mandatory session-start rules that force Claude to read state files before responding, and a /handoff skill that generates context for the next tab/session. This costs 10-15% of every session in overhead. @yurukusa <https://github.com/yurukusa>'s SessionStart hook approach (inject git log --oneline -10 at session start) is the lightest-weight version that delivers real value. But even with all of these workarounds stacked, Claude still re-discovers known issues and re-implements committed code because the memory system is append-only and unstructured — there's no weighting of facts by confidence, no deduplication, and no way to mark something as "definitively resolved." This should be a first-party feature, not a community duct-tape stack. — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/39961?email_source=notifications&email_token=AAHLVHKQDXQ7N35QQKNRSAD4TTDUDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJWHAZDKMJWHAZKM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4168251682>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHLVHNZOOYBKLDZC2TXGTT4TTDUDAVCNFSM6AAAAACXCMFPE6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCNRYGI2TCNRYGI> . You are receiving this because you were mentioned.Message ID: @.***>

junaidtitan · 4 months ago

"AI Groundhog Day" — redoing completed work across sessions — is exactly what Cozempic v1.6.11's behavioral digest + team checkpoint solve. Corrections and task state persist to disk, survive compaction, and get re-injected on session start + every 25 tool calls.

pip install cozempic && cozempic init

prisant · 4 months ago

This sounds great. Can you provide more information on what this does and
how it works?

On Mon, Apr 6, 2026 at 3:43 PM Junaid Q @.***> wrote:

junaidtitan left a comment (anthropics/claude-code#39961) <https://github.com/anthropics/claude-code/issues/39961#issuecomment-4194593109> "AI Groundhog Day" — redoing completed work across sessions — is exactly what Cozempic <https://github.com/Ruya-AI/cozempic> v1.6.11's behavioral digest + team checkpoint solve. Corrections and task state persist to disk, survive compaction, and get re-injected on session start + every 25 tool calls. pip install cozempic && cozempic init — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/39961#issuecomment-4194593109>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHLVHJGLVDSVB5QPVDJ2T34UQCGBAVCNFSM6AAAAACXCMFPE6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCOJUGU4TGMJQHE> . You are receiving this because you were mentioned.Message ID: @.***>

github-actions[bot] · 3 months ago

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

prisant · 3 months ago

Still a major problem.

On Mon, May 18, 2026 at 6:29 PM github-actions[bot] <
@.***> wrote:

github-actions[bot] left a comment (anthropics/claude-code#39961) <https://github.com/anthropics/claude-code/issues/39961#issuecomment-4482793554> Closing for now — inactive for too long. Please open a new issue <https://github.com/anthropics/claude-code/issues/new/choose> if this is still relevant. — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/39961?email_source=notifications&email_token=AAHLVHOETCJJUXLD2QXYPKD43OFFHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINBYGI3TSMZVGU2KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4482793554>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHLVHOWT33FJ7TQE6EJC5D43OFFHAVCNFSM6AAAAACXCMFPE6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DIOBSG44TGNJVGQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you were mentioned.Message ID: @.***>

github-actions[bot] · 1 month ago

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.

Session continuity: agent repeatedly redoes completed work across sessions ("AI Groundhog Day")

Summary

Evidence from real usage (2+ weeks, 7+ sessions)

The problem at each level

1. No session-to-session state

2. Plans don't track completion

3. Promises are not verified

4. Within-session memory degrades too

Workarounds I've implemented (with the agent's help)

What would actually help

Impact

12 Comments

Related Claude Code issues

Browse by topic