diff --git a/memory/strategy.md b/memory/strategy.md
new file mode 100644
index 0000000..e2f377f
--- /dev/null
+++ b/memory/strategy.md
@@ -0,0 +1,63 @@
+---
+description: Strategy for when to accumulate vs replace memory content
+---
+
+# Memory Strategy: Accumulation vs Replacement
+
+## Principle
+
+**Replacement causes collapse. Accumulation prevents it.**
+
+From the Nature 2024 paper on model collapse: training on recursively generated data causes progressive loss of signal. The same applies to memory systems — overwriting observations with summaries loses the raw data that made the summary possible.
+
+## Accumulation (Preferred)
+
+When to add new content rather than replace:
+
+- **Raw observations**: First-party observations should be preserved. Don't compress "I noticed X, Y, Z" into "I noticed patterns."
+- **User feedback**: Direct quotes and corrections are signal. Summaries lose tone and context.
+- **Unstructured knowledge**: Hypotheses, intuitions, open questions. These become more valuable over time.
+- **Low confidence additions**: If you're not sure, append rather than replace. Let the record accumulate.
+
+## Replacement (Justified)
+
+When replacing is acceptable:
+
+- **Explicit corrections**: User says "actually X" — replace the wrong belief
+- **Compression of redundant content**: If 5 observations say the same thing, synthesize into one rule + link to originals
+- **Stale information**: Config that's no longer accurate. But link to the old version so it's not lost.
+- **Security/privacy**: Redacting secrets is replacing, but justified.
+
+## Synthesis Pattern
+
+When you need to compress:
+
+1. Create a new memory block with the synthesis
+2. Keep the original blocks, but link them: "Synthesized from [original]"
+3. The synthesis is a *new* layer, not a replacement
+
+Example:
+```
+memory/observations/2026-03-27.md  # Raw observations
+memory/learned-behaviors.md        # Synthesized rules
+```
+
+The learned behaviors reference the observations. Neither replaces the other.
+
+## Implementation
+
+The `memory-verify.py` hook warns on:
+- Content shrinking > 50% (distribution narrowing)
+- Lines > 100 chars being removed (significant deletion)
+
+When you see the warning, ask:
+- Is the lost content archived elsewhere?
+- Is this a correction or a compression?
+- Would a future agent be able to reconstruct what I'm deleting?
+
+If the answer is "no" to any of these, consider accumulation instead.
+
+## References
+
+- Model Collapse paper (Nature 2024): https://www.nature.com/articles/s41586-024-07566-y
+- Issue #56: Self-improvement verification