diff --git a/data-raw/plan-v2.md b/data-raw/plan-v2.md
new file mode 100644
index 0000000..eeec274
--- /dev/null
+++ b/data-raw/plan-v2.md
@@ -0,0 +1,223 @@
+# Plan: Profile Data Structure v2
+
+## Motivation
+
+The current profile data structure (v1) was designed before `dm` existed and has several limitations:
+
+1. **Pre-aggregation**: Identical stack traces are collapsed with a `value` count, losing individual sample identity. This prevents attaching per-sample metadata (timestamps, memory snapshots).
+2. **Nested list columns**: `samples$locations` stores a list of tibbles, which complicates querying and is not representable in a relational model without unnesting.
+3. **No memory profiling support**: The format has no type-stable way to store optional memory profiling data alongside time samples.
+4. **No provenance**: There is no way to identify the source of profiling data, making it impossible to combine profiles from multiple runs while tracking origin.
+5. **Container**: The primary container is a named list (`profile_data`), rather than a `dm` object with built-in constraint checking.
+
+## Requirements
+
+1. Store raw data faithfully, without pre-aggregation
+2. Allow optional memory profiling alongside time profiling, in a type-stable way (long form)
+3. Use `dm` as the primary container
+4. Store the source/provenance of profiling data
+5. Versioned format with conversion from v1 (legacy)
+
+## Current Data Model (v1)
+
+```
+profile_data (named list of tibbles)
+├── meta:         key (chr), value (chr)
+├── sample_types: type (chr), unit (chr)
+├── samples:      value (int), locations (list of tibbles with location_id)
+├── locations:    location_id (int), function_id (int), line (int)
+└── functions:    function_id (int), name (chr), system_name (chr), filename (chr), start_line (int)
+```
+
+Key issues:
+- `samples$locations` is a nested list column — not relational
+- `samples$value` aggregates identical traces — loses individual samples
+- No provenance tracking
+- `dm_from_profile()` already unnests into a bridge table (`samples_locations`) — this should be the canonical form
+
+## Proposed Data Model (v2)
+
+The new data structure uses a `dm` object as the primary container with six tables:
+
+```
+dm object
+├── meta:              key (chr), value (chr)
+├── sources:           source_id (int), source_type (chr), source_uri (chr), source_timestamp (dbl)
+├── samples:           sample_id (int), source_id (int)
+├── sample_values:     sample_id (int), type (chr), unit (chr), value (dbl)
+├── sample_locations:  sample_id (int), depth (int), location_id (int)
+├── locations:         location_id (int), function_id (int), line (int)
+└── functions:         function_id (int), name (chr), system_name (chr), filename (chr), start_line (int)
+```
+
+### Table descriptions
+
+#### `meta`
+
+Unchanged from v1, except the version value becomes `"2.0"`.
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `key`  | chr  | Metadata key |
+| `value`| chr  | Metadata value |
+
+Required row: `key = "version"`, `value = "2.0"`.
+
+#### `sources`
+
+New table. Tracks the provenance of profiling data so that multiple profiles can be combined.
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `source_id` | int | Primary key |
+| `source_type` | chr | Origin type: `"rprof"`, `"pprof"`, `"manual"`, etc. |
+| `source_uri` | chr | File path, URL, or identifier of the source |
+| `source_timestamp` | dbl | Epoch timestamp (seconds since 1970-01-01 UTC) when the profile was captured, `NA` if unknown |
+
+Primary key: `source_id`.
+
+#### `samples`
+
+One row per raw sample (no aggregation). Each sample is linked to a source.
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `sample_id` | int | Primary key |
+| `source_id` | int | Foreign key → `sources` |
+
+Primary key: `sample_id`.
+Foreign key: `source_id` → `sources.source_id`.
+
+#### `sample_values`
+
+Long-form table for sample measurements. This is the type-stable way to store both time and memory profiling data.
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `sample_id` | int | Foreign key → `samples` |
+| `type` | chr | Measurement type: `"samples"`, `"alloc_size"`, `"dealloc_size"`, etc. |
+| `unit` | chr | Measurement unit: `"count"`, `"bytes"`, etc. |
+| `value` | dbl | The measured value |
+
+Foreign key: `sample_id` → `samples.sample_id`.
+Compound key: (`sample_id`, `type`).
+
+This replaces the separate `sample_types` table and the `samples$value` column from v1. Each sample can have multiple measurements (e.g., one row for the time sample count and one row for allocated bytes), enabling optional memory profiling in a type-stable long form.
+
+#### `sample_locations`
+
+Bridge table linking samples to their stack trace locations. Replaces the nested list column `samples$locations` from v1.
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `sample_id` | int | Foreign key → `samples` |
+| `depth` | int | Position in the stack trace (1 = innermost) |
+| `location_id` | int | Foreign key → `locations` |
+
+Foreign key: `sample_id` → `samples.sample_id`.
+Foreign key: `location_id` → `locations.location_id`.
+Compound key: (`sample_id`, `depth`).
+
+#### `locations`
+
+Unchanged from v1.
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `location_id` | int | Primary key |
+| `function_id` | int | Foreign key → `functions`, `NA` allowed |
+| `line` | int | Source line number, 0 if unknown, `NA` allowed |
+
+#### `functions`
+
+Unchanged from v1.
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `function_id` | int | Primary key |
+| `name` | chr | Demangled function name |
+| `system_name` | chr | Mangled/raw function name |
+| `filename` | chr | Source file name |
+| `start_line` | int | Start line in source file, 0 if unknown |
+
+### Relational diagram
+
+```
+sources 1──* samples 1──* sample_values
+                     1──* sample_locations *──1 locations *──1 functions
+```
+
+### Key design decisions
+
+1. **No pre-aggregation**: Each sample gets its own row in `samples`. Aggregation (run-length encoding of identical traces) is a downstream concern for analysis, not storage.
+
+2. **Long-form sample values**: The `sample_values` table stores measurements in long form. A time-only profile has one row per sample; a profile with memory data has multiple rows per sample. This is type-stable (always the same columns) and extensible (new measurement types require no schema changes).
+
+3. **Bridge table for locations**: The `sample_locations` table with a `depth` column replaces the nested list column approach. This is fully relational and works naturally with `dm`.
+
+4. **Provenance via `sources`**: Each sample is linked to a `source`, enabling combination of multiple profiles. The `source_type` and `source_uri` columns provide traceability.
+
+5. **`dm` as primary container**: The `dm` object is the canonical format, with primary and foreign keys defined. The `profile_data` class wraps or is replaced by `dm`.
+
+## Migration: v1 → v2
+
+### Conversion function
+
+`profile_v2_from_v1(x)` converts a v1 `profile_data` object to a v2 `dm` object:
+
+1. Create a single `sources` row with `source_type` inferred from the hidden `.rprof` or `.msg` component, `source_uri = NA` (not stored in v1).
+2. Expand `samples`: unnest the `value` column by repeating each sample `value` times, assigning unique `sample_id`s.
+3. Create `sample_values`: one row per sample with `type = "samples"`, `unit = "count"`, `value = 1`.
+4. Create `sample_locations`: unnest `samples$locations`, adding `depth` based on row position within each sample.
+5. Copy `locations` and `functions` unchanged.
+6. Set `meta$value` to `"2.0"`.
+
+### Version detection
+
+`validate_profile()` inspects `meta$version` to determine which validation rules to apply. The v1 validation remains for backward compatibility.
+
+### Backward compatibility
+
+- `read_rprof()` and `read_pprof()` return v2 format by default, with an option to return v1 for backward compatibility.
+- `write_rprof()` and `write_pprof()` accept both v1 and v2 formats.
+- `dm_from_profile()` is retained for v1 objects; v2 objects are already `dm` objects.
+
+## Implementation plan
+
+### Phase 1: Foundation
+
+- [ ] Add `dm` to `Imports` (currently in `Suggests`)
+- [ ] Define `new_profile_v2()` constructor that creates a `dm` with all keys
+- [ ] Define `validate_profile_v2()` validation for the v2 format
+- [ ] Implement `profile_v2_from_v1()` conversion function
+
+### Phase 2: Reader updates
+
+- [ ] Update `rprof_to_ds()` to produce v2 format directly
+- [ ] Update `msg_to_ds()` to produce v2 format directly
+- [ ] Add `source_uri` parameter to `read_rprof()` and `read_pprof()`
+- [ ] Store provenance in the `sources` table
+
+### Phase 3: Writer updates
+
+- [ ] Update `write_rprof()` to accept v2 format
+- [ ] Update `write_pprof()` to accept v2 format
+- [ ] Handle aggregation (collapsing identical traces) in the writer layer
+
+### Phase 4: Memory profiling support
+
+- [ ] Support reading Rprof memory profiling data into `sample_values`
+- [ ] Support writing memory profiling data to pprof format
+
+### Phase 5: Profile combination
+
+- [ ] Implement `combine_profiles()` to merge multiple v2 profiles
+- [ ] Ensure `source_id` and other IDs are remapped to avoid conflicts
+- [ ] Preserve provenance across combinations
+
+## Open questions
+
+1. Should `dm` move from `Suggests` to `Imports`? This adds a dependency but makes the v2 format first-class. Alternative: keep `dm` in `Suggests` and use a plain list internally, constructing `dm` on demand.
+2. Should the `profile_data` class be retained as a wrapper around `dm`, or should v2 objects simply be `dm` objects with a subclass?
+3. Should `sample_values` use `dbl` for `value` to accommodate both counts and byte sizes, or should it be polymorphic (e.g., a `vctrs` type)?
+4. Should `depth` in `sample_locations` be 1-indexed (R convention) or 0-indexed (pprof convention)?