_enhance_context_with_memories refreshes the system message by calling
MEMORY_TAG_PATTERN.sub(tagged_memory, existing_content). Passing the
memory text as a string replacement makes re.sub parse it as a template,
so backslash sequences in ordinary memories are expanded rather than
inserted.
Two failures follow, both on the second and later turns, once a block
exists to replace:
"User's repo is at C:\Users\alice" -> re.error: bad escape \U
"User writes \1 for a capture group" -> re.error: invalid group reference 1
The exception propagates out of context enhancement and aborts the turn.
A callable replacement inserts the text verbatim and cannot be parsed as
a template.
Adds a regression test covering the Windows-path case, which fails on
main with the production error.
## Stack Context
Part 2 of a 3-PR stack moving memory deduplication into the SDKs. See `sdk-dedup/tools-ts` (parent) for the full context and the TypeScript implementation this mirrors.
## What?
Port the normalized, priority-ordered (`static > dynamic > search`) profile deduplication into the Python SDKs.
- Each request injects one **owned memory block that replaces** the prior block rather than accumulating.
- Dedup is **request-local** (no shared state), so it stays correct under concurrency.
Covers OpenAI, Agent Framework (middleware + context provider), Cartesia, and Pipecat.
## Why?
Keeps the Python SDKs at behavioral parity with the TypeScript SDK so all integrations deduplicate memory the same way.
## Testing
- OpenAI: 31 passed, 11 skipped (live)
- Agent Framework: 59 passed
- Cartesia: 8 passed
- Pipecat: 8 passed
🤖 Generated with [Claude Code](https://claude.com/claude-code)
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> **Medium Risk**
> Changes memory formatting and system-prompt injection across multiple SDK integrations; incorrect dedup or replacement could alter LLM context, but there is no auth or data-store risk.
>
> **Overview**
> Ports **normalized cross-source memory deduplication** and **replace-not-append injection** into the Python OpenAI, Agent Framework, Cartesia, and Pipecat packages so they match the TypeScript SDK behavior.
>
> **Deduplication** uses request-local keys: strip optional `[YYYY-MM-DD]` prefixes, normalize whitespace, and compare with `casefold`, with priority **static → dynamic → search**. In **`query` mode**, profile static/dynamic are excluded from dedup input so facts that only appear in search (or overlap profile) are not dropped before formatting.
>
> **Injection** no longer appends memory text every turn. OpenAI and Agent Framework middleware **strip prior owned `<supermemory context="user-memories" readonly>` blocks** and **replace** them once per request while keeping the caller’s system instructions; extra system messages lose stale blocks only. New helpers (`strip`/`replace`/`wrap`) live in each package’s utils.
>
> Tests cover normalized fact variants, query-mode search retention, and stale block replacement.
>
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit 42f308b224. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
#### RE-RAISING Pipecat live speech PR
### Added native speech-to-speech model support
### Summary:
- Speech-to-speech support - Auto-detect audio frames and inject memories to system prompt for native audio models (Gemini Live, etc.)
- Fix memory bloating - Replace memories each turn using XML tags instead of accumulating
- Add temporal context - Show recency on search results ([2d ago], [15 Jan])
- New inject_mode param - auto (default), system, or user
### Docs update
- Update the docs for native speech-2-speech models
