Disclosure: I work on MemorySync (I'm Mohammed, the founder). I drafted this page with AI assistance and checked it against the source issue, the linked PR and the MemorySync LlamaIndex guide on Oct 28, 2026. I didn't run any code for this page.
The question
run-llama/llama_index#23090 reports that when memory blocks plus short-term chat history exceed token_limit, Memory._truncate_memory_blocks drops blocks in ascending priority order. So the block with priority=1 (documented as highest priority) is truncated first, and priority=2 survives. With the docs' own example (FactExtractionMemoryBlock at 1, VectorMemoryBlock at 2), condensed facts go before bulk retrieval.
Status (checked 2026-10-28): open, 3 comments. A commenter pointed to PR #23093 ("fix(core): truncate lowest-priority memory blocks first"), but that PR is now closed without being merged, so no fix has landed as far as the issue shows.
Workarounds
Not run for this page; the issue includes a short repro you can run against your installed version first.
-
Use
priority=0for the block you can't lose. Per the field's docs, 0 means never truncate; the issue doesn't report that being broken. -
Raise
token_limitor shrink blocks so truncation doesn't trigger in normal conversations. - Don't rely on the 1-vs-2 order until a fix ships. If you must, test which block survives on your version with the issue's repro and set priorities to match the observed behaviour, then re-check after upgrades.
Common questions
Is there a fix?
Not as of Oct 28, 2026: the issue is open, and PR #23093 was closed without being merged.
Which priority is safe?
priority=0, documented as never truncated. The issue doesn't report that being broken.
Where MemorySync fits
Only if you use it already or are evaluating it. The MemorySync LlamaIndex guide documents a MemorySyncMemoryBlock whose atruncate drops its lowest-ranked lines instead of the whole block. That changes what happens to that one block when it's truncated; it doesn't fix the order bug in Memory itself. To see save and recall in your browser first, try the MemorySync playground.
Sources
- run-llama/llama_index#23090, "[Bug]: Memory truncates highest-priority memory blocks first" (read 2026-10-28)
- run-llama/llama_index PR #23093 (closed, not merged; read 2026-10-28)
- MemorySync LlamaIndex guide: docs.memorysync.io/guides/llamaindex
Top comments (0)